Journal of Communication and Computer Volume 10, Number 12, December 2013 (Serial Number 109)
David Publishing
David Publishing Company www.davidpublishing.com
Publication Information: Journal of Communication and Computer is published monthly in hard copy (ISSN 1548-7709) and online (ISSN 1930-1553) by David Publishing Company located at 240 Nagle Avenue #15C, New York, NY 10034, USA. Aims and Scope: Journal of Communication and Computer, a monthly professional academic journal, covers all sorts of researches on Theoretical Computer Science, Network and Information Technology, Communication and Information Processing, Electronic Engineering as well as other issues. Contributing Editors: YANG Chun-lai, male, Ph.D. of Boston College (1998), Senior System Analyst of Technology Division, Chicago Mercantile Exchange. DUAN Xiao-xia, female, Master of Information and Communications of Tokyo Metropolitan University, Chairman of Phonamic Technology Ltd. (Chengdu, China). Editors: Cecily Z., Lily L., Ken S., Gavin D., Jim Q., Jimmy W., Hiller H., Martina M., Susan H., Jane C., Betty Z., Gloria G., Stella H., Clio Y., Grace P., Caroline L., Alina Y.. Manuscripts and correspondence are invited for publication. You can submit your papers via Web Submission, or E-mail to informatics@davidpublishing.org. Submission guidelines and Web Submission system are available at http://www.davidpublishing.org, www.davidpublishing.com. Editorial Office: 240 Nagle Avenue #15C, New York, NY 10034, USA Tel：1-323-984-7526, Fax: 1-323-984-7374 E-mail: informatics@davidpublishing.org; com58puter@hotmail.com Copyright©2013 by David Publishing Company and individual contributors. All rights reserved. David Publishing Company holds the exclusive copyright of all the contents of this journal. In accordance with the international convention, no part of this journal may be reproduced or transmitted by any media or publishing organs (including various websites) without the written permission of the copyright holder. Otherwise, any conduct would be considered as the violation of the copyright. The contents of this journal are available for any citation. However, all the citations should be clearly indicated with the title of this journal, serial number and the name of the author. Abstracted / Indexed in: Database of EBSCO, Massachusetts, USA Chinese Database of CEPS, Airiti Inc. & OCLC Chinese Scientific Journals Database, VIP Corporation, Chongqing, P.R.China CSA Technology Research Database Ulrich’s Periodicals Directory Summon Serials Solutions Subscription Information: Price (per year): Print $520; Online $360; Print and Online $680 David Publishing Company 240 Nagle Avenue #15C, New York, NY 10034, USA Tel：1-323-984-7526, Fax: 1-323-984-7374 E-mail: order@davidpublishing.org
D
DAVID PUBLISHING
David Publishing Company www.davidpublishing.com
Journal of
Communication and Computer Volume 10, Number 12, December 2013 (Serial Number 109)
Contents Computer Theory and Computational Science 1465
Ranking Approach for the User Story Prioritization Methods Sultan Alshehri and Luigi Benedicenti
1475
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia Mohamed Fathi Alaweti, Nurdiana Azizan and Qais Faryadi
1484
Optimal Object Categorization under Application Specific Conditions Steven Puttemans and Toon Goedeme
1497
Image Super Resolution and Enhancement Using E-spline Gamal Fahmy
Network and Information Technology 1502
New Cloud Consolidation Architecture for Electrical Energy Consumption Management Nawfal Madani, Adil Lebbat, Saida Tallal and Hicham Medromi
1507
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real Indoor and Outdoor Location in WSN Mauro Rodrigo Larrat Frota e Silva, Leomรกrio Silva Machado and Dionne Cavalcante Monteiro
1522
Performance Evaluation of Quicksort with GPU Dynamic Parallelism for Gene-Expression Quantile Normalization Roberto Pinto Souto, Carla Osthoff, Douglas Augusto, Oswaldo Trelles and Ana Tereza Ribeiro de Vasconcelos
1529
A Privacy Taxonomy for the Management of Ubiquitous Environments Valderi Reis Quietinho Leithardt, Guilherme Antonio. Borges, Anubis Graciela de Morais Rossetto, Carlos Oberdan Rolim, Claudio Fernando Resin Geyer, Luiz Henrique Andrade Correia, David Nunes and Jorge Sa Silva
Communications and Electronic Engineering 1554
Triangle Routing Problem in Mobile IP Sherif Kamel Hussein and Khaled Mohamed ALmustafa
1566
Geographical Monitoring of Electrical Energy Quality Determination: The Problems of the Sensors Maurizio Caciotta, Fabio Leccese, Sabino Giarnetti and Stefano Di Pasquale
1573
Controller Design for Networked Control System with Uncertain Parameters Caixia Guo, Xuebing Wu and Xinwei Yang
Journal of Communication and Computer 10 (2013) 1465-1474
Ranking Approach for the User Story Prioritization Methods Sultan Alshehri and Luigi Benedicenti Software Systems Engineering, University of Regina, Regina SK S4S 6W2, Canada
Received: November 11, 2013 / Accepted: December 12, 2013 / Published: December 31, 2013. Abstract: The AHP (analytic hierarchy process) has been applied in many fields and especially to complex engineering problems and applications. AHP is capable of structuring decision problems and finding mathematically determined judgments built on knowledge and experience. This suggests that AHP should prove useful in agile software development, where complex decisions occur routinely. This paper describes a ranking approach to help stakeholders select the best prioritization method for prioritizing the user stories. Key words: Extreme programming, user stories, analytic hierarchy process.
1. Introductionď€ The quality of Extreme Programming (XP) development results from taking 12 core practices to their logical extremes [1]. One such practice is the planning game, in which customers and developers cooperate to develop requirements that produce the highest value for customers as rapidly as possible. This is accomplished as follows. Customers write system requirements as user stories. User stories are defined as “short descriptions of functionality told from the perspective of a user that are valuable to either a user of the software or the customer of the softwareâ€? [2]. Developers review the stories to ensure domain-specific information is sufficient for their implementation. Developers evaluate user stories using story points to identify the complexity and cost of their implementation. Then, user stories are broken down into small tasks. Finally, customers and developers collaborate in prioritizing user stories based on their value and other relevant factors. To reconcile conflicting opinions among them, customers and developers often adopt a prioritization Corresponding author: Sultan Alshehri, Ph.D., research field: agile methodology. E-mail: lowtdree2@hotmail.com.
method [3-5]; but this adoption process is usually not formalized. In this paper, the AHP (analytical hierarchy process) is utilized as a well-structured multi-criteria decision making tool to help XP Software development teams rank six prioritization methods: 100-Dollar test (or cumulative voting), MoSCow, top-ten Requirements, Kano model, theme screening, and relative weighting. This paper is organized as follows: Section 2 briefly explains the AHP method; the six prioritization methods are presented in Section 3; four criteria for ranking the prioritization methods are proposed in Section 4; a case study, its results and its findings are presented in Section 5, and Section 6 concludes the paper.
2. Related Work There is no consensus in the literature on the most important factors determining the priority of system requirements. However, almost all the factors taken into consideration aim to maximize the value delivered to the customer. Bakalova et al. [6] proposed to use project context, effort estimation, dependencies, input from the developers, learning experiences and external change. Hoff et al. [7] relied on four factors:
1466
Ranking Approach for the User Story Prioritization Methods
cost-benefit to the organization, impact of maintenance, complexity and performance effects. They also considered fixed errors, requirement dependencies, complexity, and delivery data/schedule as ancillary factors. Somerville and Sawyer [8] prioritize requirements based on the viewpoint approach that represents information about the system requirements from different perspectives representing different types of stakeholder. Davis used Triage [9] as an evaluation process considering time, available resources, and requirements interdependencies. Lutowski [10] prioritized the requirements based on the importance or immediacy of need. Bhoem [11] considered the cost of implementing the requirement as the most important factor for prioritization. In Bhoem’s work, cost is related to the technical environment, complexity, quality, timeframe, documentation, availability reusable software, participant competencies, and stability of requirements. Berander and Andrews [12] surveyed the literature and found common aspects in prioritizing requirements such as penalty, cost, time, risk and volatility. The authors added that other aspects like financial benefits, competitors, release theme, strategic benefit, competence/resource, and ability to sell should also be considered. In the agile methodology domain, Patel and Ramachandran [13] prioritized user stories based on business functionality, customer priority, core value, market values, implementation cost, and business risk. Many well-established prioritization methods available are applicable to requirements prioritization: Ping Pong Ball, Pair-Wise Analysis, Weighted Criteria Analysis, Dot Voting, Binary Search Tree, Ranking, Numeral Assignment Method, Requirements Triage, Wieger’s Matrix Approach, Quality Function Deployment, Bucket Method, Cumulative Voting, Round-the-Group Prioritization, Theory-W, and Theme Scoring [14, 15]. Changes to requirements in a plan-based environment are difficult and costly. Thus, a change
the user considering simple may translate into a painful process for the developers. By definition, this is not the case for requirements in agile methods. This fundamental difference may have an impact on the optimal choice of prioritization method. Mead conducted a case study to determine the most suitable requirements prioritization methods to be used in software development [5]. This study compared three common methods: Numeral Assignment Method, Theory-W, and AHP. The prioritization method comparison was based on five aspects: clear-cut steps, quantitative measurement, high maturity, low labor-intensity, and shallow learning curve. The results indicated that the AHP ranked the highest score of 16, while the Numeral Assignment Method scored a 12, and Theory-W scored an 8.
3. Methodology The primary objective of this study is to investigate how the AHP can be used to rank the user stories prioritization methods. The methodology used in this study is the case study methodology described in Ref. [16]. The following research questions provided a focus for our case study investigation: (1) How does the AHP help select a prioritization method for user stories? (2) How do the AHP results affect the relationships among developer’s relation and their performance? The units of analysis for this study derive from these research questions. The main focus is to rank several tools that can be used to prioritize user stories. Accordingly, ranking and the evaluation process are two the units of analysis for this study. Also, we consider the developers view of how the AHP benefits each XP practice. As result, our study is designed as multiple cases (embedded) with two units of analysis.
4. Data Collection and Sources In the beginning of the study, we found the criteria
Ranking Approach for the User Story Prioritization Methods
1467
affecting the ranking process and helping to examine the AHP tool ability and benefits. This data was collected from literature review and previous studies. To increase the validity of this study, data triangulation was employed. The data sources in this study were: Archival records such as study plans from the graduate students; Questionnaire given to the participants when developing the XP project; Open-ended interviews with the participants; Feedback from the customer.
foundation specifically in the iteration plan, release planning and prioritizing the user stories. In addition, the students were exposed to the AHP methodology and learned the processes necessary to conduct the pairwise comparisons and to do the calculations. Several papers and different materials about AHP and user stories were given to the students to train them and increase their skills in implementing the methodology. Finally, a survey was distributed among students to get further information about their personal experiences and knowledge.
5. Case Study
6. The AHP
The case study was conducted in the Advanced Software Design Course offered to graduate students in fall 2012 at the University of Regina. The participants were 12 master’s students and a client from a local company in Regina. Participants had various levels of programming experience and a good familiarity with XP and its practices. The student’s background relevant for the case study included several programming languages such as Java, C, C#, and ASP.net. All participants had previous project development experience. The study was carried out throughout 15 weeks; the students were divided into two teams. Both teams were assigned to build a project called “Issue Tracking System” brought in by the client along with a set of requirements compatible with current industry needs. The project evolved through 5 main iterations and by the end of the semester, all software requirements were implemented. The students were requested to try all requirements in each prioritization method before applying AHP to rank them. Participants were given detailed lectures and supporting study materials on extreme programming practices that focused on planning game activities which included writing user stories, prioritizing the stories, estimating process parameters, and demonstrating developers’ commitments. The students were not new to the concept of XP, but they gained more knowledge and
AHP is a systematic approach for decision-making that involves the consideration of multiple criteria by structuring them in a hierarchical model. AHP reflects human thinking by grouping the elements of a problem requiring complex and multi-aspect decisions [17]. The approach was developed by Thomas Saaty [8] as a means of finding an effective and powerful methodology that can deal with complex decision-making problems. AHP comprises the following steps: (1) Structure the hierarchy model for the problem by breaking it down into a hierarchy of interrelated decision elements. (2) Define the criteria or factors and construct a pairwise comparison matrix for them; each criterion on the same level of the decision hierarchy is compared with other criteria in respect of their importance to the main goal. (3) Construct a pairwise comparison matrix for alternatives with respect to each objective in separate matrices. (4) Check the consistency of the judgment errors by calculating the consistency ratio. (5) Calculate the weighted average rating for each decision alternative and choose the one with the highest score. More details on the method, including a step-by-step example calculation, are found in Ref. [17]. Saaty [18] developed a numerical scale for assigning the weight for criteria and alternative by giving a value between 1 (equal importance) and 9 (extreme importance); see Table 1 for details.
1468
Ranking Approach for the User Story Prioritization Methods
Table 1 AHP numerical scale developed by Saaty. Scale Equal importance Moderate importance of one over other Very strong or demonstrated importance Extreme importance Intermediate values
Numerical rating 1
1
3
1/3
7
1/7
9 2, 4, 6, 8
1/9 1/2, 1/4, 1/6, 1/8
Reciprocal
7. Prioritization Methods There are several methods for prioritizing system requirements; the six most commonly used can be summarized as follows: 7.1 The 100-Dollar Test (Cumulative Voting) This is a straightforward method described by Leffingwell and Widrig [19] where each stakeholder gets 100 imaginary units (money, hours, etc.) to distribute among the given requirements. If the requirements are too many, it is recommended to use more units of value for more freedom in the prioritization [20]. After distributing the units on the requirements, stakeholders calculate the total for each requirement and rank the requirements accordingly.
problem with this method is the difficulty of distinguishing the terms “Must” and “Should” as they both express a customer preference or desire. 7.3 Top-Ten Requirements In this approach, the stakeholders select their Top-Ten Requirements without giving them a specific priority [23]. This is to avoid the conflict between stakeholders that may arise from the desire to support specific
requirements.
However,
if
stakeholder
alignment is low, it is possible that none of the choices for some stakeholders will appear in the aggregated top priority requirement list. 7.4 Kano Model This
method
was
established
for
product
development by Noriako Kano [24] to classify the requirements into five categories based on the answers to two questions about every requirement: (1) Functional question: “How do you feel if this feature is present?”; (2) Dysfunctional question: “How do you feel if this feature is not present?”. The customer has to choose one of the five following options for the answers [25]:
7.2 MoSCoW
(1) I like it.
This is one of the methods for prioritization originating from the DSDM (dynamic software development method) [21]. The requirements are classified into four groups depending on the importance of the functional requirements [22]: M: MUST have this requirement. It is the highest priority and without it the project is considered a failure. S: SHOULD have this requirement if possible. Customer satisfaction depends on this requirement. But we cannot say its absence causes a project to fail. C: COULD have this requirement if it does not affect anything else. W: WON’T have the requirement this time but WOULD like to in the future. This method helps understand customer needs. The
(2) I expect it. (3) I’m neutral. (4) I can tolerate it. (5) I dislike it. 7.5 Themes Screening This is a method employed when stakeholders have many relevant user stories that need to be grouped together. While writing the stories, stakeholders eliminate similar stories or ones that have already been covered by others. Then they follow the steps below [26]: (1) Identify 5-9 (approximately) selection criteria that are important in prioritizing the themes. (2) Identify a baseline that is approved and understood by all the team members.
Ranking Ap pproach for the t User Storry Prioritizatio on Methods
(3) Comppare each theeme to the baaseline themee for each criterioon. Use “+” for themes that t rank “beetter than” the baseline them me, “–” for themes t that rank r “worse thann” the baselinne theme andd “0” for theemes that rank “eqqual” to the baseline b them me. (4) Calcuulate the “net score” by sum mming up alll the plusses and minuses. Raank as numbeer one the thheme that receivedd the highest net score. 7.6 Relative Weighting e This metthod involvves the evalluation of each requirement based on thee effect of itss presence and its absence. A scale from 0 to 9 is ideentified for each e requirement, 0 being a low effect and 9 being a high h effect. Stakeeholder will give g every feeature a valuee for its presence as well as a penalty forr its absence and estimate itss implementtation cost. The priorityy is calculated by dividing thhe total value by the total cost to generate a prioritizatioon indicator [226].
8. Proposeed Criteria for Rankin ng To rank eaach method, it i is necessaryy to determinee the criteria thatt affect stakkeholders whhen choosinng a prioritizationn process. Thhese prioritizaation criteria will then be com mpared amoong each othher. Finally, the prioritizationn methods will w be compaared against each e of the criterria [27]. In this paper, we w propose four prioritizationn criteria thatt emerged durring the coursse of the case studdy we conduccted, but the method m descrribed in this paperr can be appllied to any seet of criteria. The criteria show wn below arre simply illustrative off the prioritizationn method. (1) Simplicity: What is the simpllest prioritizaation
Fig. 1 AHP structure for ranking r the prrioritization meethods.
14699
metthod in term ms of easee of understtanding andd app plication? (2 2) Time: Whhich one of thhese methods will save thee mosst time when the team appplies it to the user u stories? (3) Accuracy: Which one oof these meth hods will givee the most accuratte results? (4 4) Collaboraation: Which oone of these methods m willl ach hieve the highhest degree off collaboratio on among thee stak keholders andd the XP team m in general?
9. AHP A in Praactice The T first step in the analyttic hierarchy process is too stru ucture the prroblem as a hierarchy. In n this paper,, such a hierarchyy includes thrree levels. Th he top level iss the main objectiive: ranking tthe prioritizattion methodss. Thee second leevel is thee prioritizatiion criteria: sim mplicity, timee, accuracy, and collabo oration. Thee thirrd level is the alternatives: 100-D Dollar Test,, Top p-Ten Requuirements, Kano Mod del, Themee Scrreening, Relaative Weighting, and MoS SCow. Fig. 1 illu ustrates the AHP A hierarchhy we choossing for thiss pap per. Then, T the hierrarchy is useed to generatee appropriatee AH HP tables. Alll team membbers receive these tables,, whiich shortens the t time to fiill them and facilitates f thee com mparison proocess. A coover page dedicated d too colllecting generral informatioon of each teeam memberr inclluding experiience, type, aand level of programming p g skillls is also haanded out. A matrix is then t used too com mpare the fouur prioritizatioon criteria. Accordingly, A we requiredd all students to use thee prio oritization methods m throoughout the project too exp perience their advantages a annd disadvantag ges. Then, wee
1470
Ranking Ap pproach for the t User Storry Prioritizatio on Methods
asked the stuudents to evaaluate these methods m basedd on the prioritizzation criteriia. To accom mplish this, we provided theem with the AHP tables and cover page p described abbove. The studeents first com mpared the critteria among each e other using the Saaty scaale, ranging from f 1 to 9. The students used a check listt with the folllowing questiions: Which is more important: simplicity or time and by how mucch? Which is more important: simplicity or accuracy and by how much? Which is more important: simplicity or collaborationn and by how w much? Which is more impoortant: time orr accuracy and by how much? Which is more impportant: time or collaboraation and by how much? Which is more important: accuracy or collaborationn and by how w much? After finisshing the criteeria comparissons, the studdents had to evaluuate all the prioritization p methods agaainst each other based on eaach criterion every time. An example folllows: In term m of simpplicity, whicch is simpplest: 100-Dollar or o Top-Ten annd by how much? m The samee questions annd comparisoons were repeeated for all prioriitization methhods and criteeria.
(Fig g. 2). It appeaars that accuracy was the most m relevantt criterion for the t team, ffollowed by y simplicity,, colllaboration and time. The T results foor Team 2 paaint a somew what differentt pictture: The Rellative Weightting method is still on top,, but it is follow wed by Topp-Ten, MoSccoW, Themee Screening, 100--Dollar and finally Kan no. Table 3 prov vides the relative r scorees of each ranking ass perccentages. As A for the impportance of each criterion as perceivedd by Team 1 (Fig. 3), it appeaars that accurracy was stilll the most relevannt prioritizatiion criterion, followed byy time, collaboratiion and simpllicity. Tab ble 2
Prioritizzation methodss ranking for Team T 1.
Meethods Rellative Weightinng Mo oScoW Theem Screening Kan no Top p-Ten 100 0-Dollar
Scores (%) 244.39 200.38 177.70 155.81 122.75 8..97
10. Findin ngs and Ressults Each s student inddividually evaluated the prioritizationn methods based onn the critteria mentioned earlier. e The Expert E Choicce Software [28] was used too calculate thhe aggregatioon results forr the two teams. The resullts for Team 1 show that the highest rank r was given too the relativee weighting method, m folloowed by MoScoW W, Theme Screening, Kano, Top--Ten Requiremennts and 100-D Dollar Test. Table T 2 provvides the relative scores s of eachh ranking as percentages. p The software also allows us to t examine the importance of each criteerion as perceeived by Teaam 1
Fig.. 2 The imporrtance of the crrtiteria by Tea am 1. Tab ble 3
Prioritizzation methodss ranking for Team T 2.
Meethods Rellative Weightinng Top p-Ten Mo oScoW Theeme Screening 100 0-Dollar Kan no
Scores (%) 322.67 266.12 155.44 155.35 7..15 3..27
Ranking Ap pproach for the t User Storry Prioritizatio on Methods
1471
t software to aggregate results from m whiich allowed the team m members. 11.2 2 Interview Results R The T interview ws were condducted after showing thee partticipants the results r of the A AHP evaluatiion for all thee XP practices. Soome of the reesults were su urprising andd ncluded openn otheers were exppected. The iinterviews in queestions to obtaain the studennts’ general op pinions aboutt AH HP, the advanttages and dissadvantage off using AHP,, and d the best exxperience of AHP among g all the XP P mportance of the t criteria by Team 2. Fig. 3 The im
11. Observvations 11.1 AHP Raanking Resultts When all the criteria were w consideered together,, the Relative Weeighting Methhod was ranked the highesst by both teams. The MoScoW W Method was w ranked inn the second posittion by Team m 1 and third position p by Team T 2. The 100-D Dollar Test Method M was ranked r in thee last position by Team T 1 and inn the second to t last position by Team 2. Both teaams considerred accuracyy as the most m important crriteria. Simpllicity in Team m 1 and tim me in Team2 respeectively weree considered to be the seccond highest impoortant criterioon. When thhe prioritizaation methodds were rannked considering each criterioon individuallly, we found that for Team 1 the MoScooW Method was ranked the highest in terms of siimplicity annd time criteeria. Relative Weeighting was ranked r the hiighest in term ms of accuracy annd collaboratiion criteria. Results R relateed to Team 2 are slightlyy different: the Top--Ten Requiremennts Method raanked the higghest in term ms of simplicity and time criteria. c Relative weighhting ranked thee highest inn terms off accuracy and collaboratioon criteria. These ressults are inddicative of different d chooices made in eachh team. Althoough the rankking was achieeved through inddividual compparisons, thee group behaavior was consisteent as reflectted in the consistency scoores,
pracctices. As nooted previouslly, the data was w collectedd in th he form of haandwritten nootes during th he interviews.. Theese notes werre organized iin a folder fo or the sake off easy y access and analysis. From F the innterviews, w we found very positivee feed dback from thhe participannts regarding AHP. It wass felt that AHP reesolved any conflicting opinions o andd brou ught each teaam member’ss voice to the decision in a pracctical way. AHP A also emphhasized the courage of thee team m by letting every opinionn be heard. The T time andd the number of coomparisons w were the main n concerns off the participants. All of them rrecommended using AHP P t future wiith XP. Therre were a few additionall in the reco ommendationns as well, such as deeveloping ann auto omated tool to reduce the ttime required d for the AHP P calcculation, addding the mobility featuress, performingg costt and risk annalysis, and trying AHP in other XP P areaas and studyinng the outcom mes. 11.3 3 Questionnaaires Questionnaire Q es were also ggiven to the participants inn order to obtain their percepptions of and d experiencess with h AHP. The questionnairres were divided into twoo maiin parts. Thee first part contained queestions aboutt AH HP as a decisiion and rankiing tool. Thee second partt con ntained questiions regardingg the direct benefits of thee XP practice and a investiggated the participants’ kert scale too satiisfaction. Wee used a sevven-point Lik refllect the levell of acceptabbility of the AHP A tool ass folllows:
1472
Ranking Approach for the User Story Prioritization Methods
(1) Totally unacceptable (2) Unacceptable. (3) Slightly unacceptable. (4) Neutral. (5) Slightly acceptable. (6) Acceptable. (7) Perfectly acceptable. Once the participants completed the questionnaire, we aggregated the responses and presented the total percentage of acceptability for each statement. The total percentage of the acceptability was calculated as follows: 11.4 The TPA (Total Percentage of Acceptability) The TPA = the average score for each team × 100 / 7. 11.5 The Average Score for Each Team The average score for each team = the sum of the scores given by the team members / number of members in the team. The following percentages show the acceptability level for the AHP as aranking tool: Improving team communication: Team 1 scored 83% and Team 2 scored 86%. Creating a healthy discussion and learning opportunities: Team 1 scored 74% and Team 2 scored 93%. Clarifying the ranking problem: Team 1 scored 86% and Team 2 scored 93%. Resolving conflicting opinions among members: Team 1 scored 78% and Team 2 scored 93%. Increasing team performance: Team 1 scored 74% and Team 2 scored 88%.
12. Validity Construct validity, Internal Validity, External Validity and Reliability categorize common threats to the validity of the study [29]. “Empirical studies in general and case studies in particular are prone to biases and validity threats that make it difficult to control the quality of the study to generalize its results”
[30]. In this section, relevant validity threats are described. A number of possible threats to the validity of this work can be identified. 12.1 Construct Validity Construct validity deals with the correct operational measures for the concept being studied and researched. The major threat to this study is the small number of participants in each case study. This threat was mitigated by using several methods in order to ensure the validity of the findings. Data triangulation: A major strength of case studies is the possibility of using many different sources of evidence [29]. This issue has been taken into account through the use of surveys and interviews with different types of participants from different environments with various levels of skills and experience, and through the use of several observations as well as feedback from those involved in the study. By establishing a chain of evidence, we were able to reach a valid conclusion. Methodological triangulation: The research methods employed were a combination of a project conducted to serve this purpose, interviews, surveys, AHP results comparisons, and researchers’ notes and observations. Member checking: Presenting the results to the people involved in the study is always recommended, especially for qualitative research. This has been done by showing the final results to all participants to ensure the accuracy of what was stated and to guard against researcher bias. 12.2 Internal Validity Internal validity is only a concern for an explanatory case study [29], and it focuses on establishing a causal relationship between students and educational constraints. This issue can be addressed by relating the research questions to the propositions and other data sources providing information regarding the questions.
Ranking Approach for the User Story Prioritization Methods
12.3 External Validity
Reference
External validity is related to the domain of the study and the possibilities of generalizing the results. To provide external validity to this study, we will need to conduct an additional case study in the industry involving experts and developers and then observe the similarities and the differences in the findings of both studies. Thus, future work will contribute to accrue external validity.
[1]
12.4 Reliability
[6]
[2] [3] [4] [5]
Reliability deals with the data collection procedure and results. Other researchers should arrive at the same case study findings and conclusions if they follow the same procedure. We address this by making the
[7]
research questions, case study set up, data collection and analysis procedure plan available for use by other [8]
researchers.
13. Conclusions [9]
After using AHP to rank the most common requirement prioritization methods used in XP development to prioritize the user stories, AHP was found to be a relevant and useful tool that affords very
[10] [11]
good vision to stakeholders when they want to decide on which prioritization method is the most suitable. Considering
simplicity,
time,
accuracy
[12]
and
collaboration when selecting a prioritization method
[13]
could bring many advantages to the XP team, including the stakeholders. The relative weighting method was the most preferred method for both teams in our case
[14]
study, but the procedure we followed is general and thus the ranking can change depending on the team. More importantly, though, AHP helped students evaluate each prioritization method from different
[15]
viewpoints. In addition, they could mathematically reconcile the conflicts of opinions among them. AHP introduces a cooperative decision making environment, which accelerates the XP development process and maximizes the effectiveness of the software developed.
[16] [17]
1473
K. Beck, Extreme Programming Explained: Embrace Change, 2nd ed., Addison-Wesley, 2000. M. Cohn, Advantage of User Stories for Requirements, Information Network, October 2004. K. Wiegers, Software Requirements, Microsoft Press, Redmond, USA, 2003. S. Lauesen, Software Requirements-Styles and Techniques, Pearson Education, Essex, ,UK, 2002. R. Mead, Requirements Prioritization Introduction, Software Engineering Institute, Carnegie Mellon University, 2006. Z. Bakalova, M. Daneva, A. Herrmann, R. Wieringa, Agile requirements prioritization: What happens in practice and what is described in literature, Requirements Engineering: Foundation for Software Quality Lecture Notes in Computer Science 6606 (2011) 181-195. G. Hoff, A. Fruhling, K. Ward, Requirements Prioritization Decision Factors for Agile Development Environments, University of Nebrask, Omaha, USA, 2008. Sommerville, P. Sawyer, Requirements Engineering: A Good Practice Guide, John Wiley & Sons Ltd, Chichester, England, 1997. Davis, The Art of Requirements Triage, IEEE Computer 36 (2003) 42- 49. R. Lutowski, Software Requirements, Auerbach Publications, Boca Raton, USA, 2005. B. Boehm, The High Cost of Software, Practical Strategies for Developing Large Software Systems, Addison-Wesley, Reading MA, 1975. P. Berander, A. Andrews, Requirement prioritization, in: Engineering and Managing Software Requirements, Berlin, Deutschland, 2005. C. Patel, M. Ramachandran, Story card based agile software development, International Journal of Hybrid Information Technology 2 (2009) 125-140. Z. Racheva, M. Daneva, L. Buglione, Supporting the dynamic reprioritization of requirements in agile development of software products, in: Second International Workshop on Software Product Management IWSPM’08, Barcelona, Catalunya, Sept. 09, 2008. Q. Ma, The effectiveness of requirements prioritization techniques for a medium to large number of requirements: A systematic literature review, Master Thesis, Auckland University of Technology, 2009. K. Yin, Case Study Research: Design and Methods, 2nd ed., SAGE Publications, 1994. N. Tiwari, Using the Analytic Hierarchy Process (AHP) to Identify Performance Scenarios for Enterprise
1474
Ranking Approach for the User Story Prioritization Methods
Applications, 2006. [18] T. Saaty, The Analytic Hierarchy Process, McGraw-Hill, New York, 1980. [19] D. Leffingwell, D. Widrig, Managing Software Requirements: A Use Case Approach, 2nd ed., Addison-Wesley, Boston, USA, 2003. [20] P. Berander, C. Wohlin, Different in views between development roles in software process improvement—A quantitative comparison, in: Proceedings of the IEEE 8th International Conference on Empirical Assessment in Software Engineering (EASE 2004), Stevenage, UK, 2004. [21] K. Waters, Prioritization Using MoSCoW, Agile Planning, January 12, 2009. [22] The MoSCoW Prioritization Technique LMR Technologies, Agile Practices: Scrum, XP, Lean, Kanban: www.lmrtechnologies.com (accessed: Jun. 22, 2012). [23] K. Wiegers, First things first: prioritizing requirements,
Software Development 7 (1999) . [24] E. Zultner, H. Mazur, The Kano Model: Recent Developments, in: The Eighth Symposium on Quality Function Deployment, Austin, 2006. [25] Hand, Applying the Kano Model to User Experience Design, UPA Boston Mini-Conference, May 2004. [26] M. Cohn, User Stories Applied for Agile Software Development, Addison-Wesley Professional, 2004. [27] T. Saaty, How to make a decision: The analytic hierarchy process, Interfaces 24 (1994) 19-43. [28] Expertchoice for Collaborative Decision Making [Online], http://www.expertchoice.com (accessed: Dec. 05, 2012). [29] R.K. Yin, Case Study Research—Design and Methods, 3rd ed., Sage Publications, London, 2003. [30] R. Lincke, How do Ph.D. Students Plan and Follow-up their Work?—A Case Study, University Sweden, April, 2007.
Journal of Communication and Computer 10 (2013) 1475-1483
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia Mohamed Fathi Alaweti, Nurdiana Azizan and Qais Faryadi Department of Science and Technology, Faculty of Science and Technology, University Sains Islam Malaysia, Nilai 71800, Negeri Sembilan, Malaysia
Received: November 09, 2013 / Accepted: December 10, 2013 / Published: December 31, 2013. Abstract: The technology of E-procurement becomes more popular in many developed countries due to its accuracy in decision making for big projects that enhance the public bidding process for development projects in any country. The experience of employees who operate the system of E-procurement and the perception of users who bid for projects are essential to ensure the efficiency of procurement and the whole bidding process for public projects. This paper examined the impact of employee’s experience on the efficiency of E-procurement systems in public agencies. In addition, we attempt to examine the perception factor of using E-procurement systems with regard to users in the seller side. A survey was conducted to identify the impact of experience and perception on the efficiency towards adoption and use of E-procurement system among users who working in firms participated in online bidding for public agencies projects, and employees who are working in E-procurement department in government administrations Malaysian government. A total of 80 questionnaires were collected and the data were analyzed to look at the level of the impact of experience and perception in E-procurement users. The general findings indicate a positive attitude experience and perception among the users in using the E-procurement. Key words: E-procurement, user’s perception, operator’s experience, E-procurement efficiency.
1. Introduction E-procurement or what is called electronic procurement is the B2B or B2C purchasing process and sale of supplies and services using the Internet technology and information systems that construct the whole E-procurement system, such as electronic data interchange and enterprise resource planning [1]. E-procurement is operated using a software application that includes specific features for supplier management and complex auction functions. The new generation of E-procurement is more custom design according to the needs of the buyer, which is called a software-as-a-service [2]. In other words, E-procurement is an Internet technology Corresponding author: Mohamed Alaweti, Ph.D. student, research fields: E-government, E-procurement system, and the transparent of E-procurement and information technology on public projects and E-commerce. E-mail: mralakrami@gmail.com.
solution facilitating corporate buying using the Internet [3]. In all developed countries, it is a substantial part of the national economy, and shifting its business activities through online biding using E-procurement has the potential to provide major impetus to the rollout of new technologies throughout the economy. However, the operational benefits of E-procurement technology for government agencies are beyond question. Installing new technology can be simple, but experience has shown that extracting maximum benefit involves governance, management, organizational and behavioral changes are almost always complex [4]. The system of E-procurement chain consists of catalogue management, E-payment, E-tendering, E-auctioning, vendor management application, purchase order integration, order status, E-invoicing, ship notice, and contract management. Since the Internet arrived on the scene as a supply
1476
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia
management tool in the mid-1990s, organizations attempted to gain the benefits of using E-procurement in bidding for big and public projects. The main advantaged of E-procurement is to both sellers and buyers are cost reduction, process streamlining, effective contracting, and effective bidding management. However, both suppliers and contractors to use E-procurement facing many challenges, and one of these challenges is the experience of IT personnel who operate the system, and also the perception of sellers to bid online for public projects. The experience of IT employees who run E-procurement applications is very important in the whole E-procurement process and their role is essential in the workflow of assessment of bids involved in the preparation of tenders. The experience is linked with the efficiency of the whole E-procurement system; therefore, providing qualified employees who are responsible for running E-procurement application effectively is an essential requirement for efficient and reliable E-procurement process. This paper assumes that two elements specify the experience factor in E-procurement systems, which are operator’s background and user’s perception to work with complicated electronic bidding systems. The background on using E-procurement systems from the side of buyer should be compliment by the experience and perception of the seller or bidders. 1.1 Operator Experience A procurement specialist or operator is the person within a company or organization who is responsible for acquiring the supplies, equipment, materials, property, and services necessary for the business’ existence. He or she works with the company’s financial department to secure the purchases, negotiate contracts with vendors, and implement cost-reducing methods. This person should be aware of the organization’s budget and strive to match his or her plans with the company’s financial situation. To become a procurement specialist, an individual should
have a background in business, marketing or law [5]. The operator who runs E-procurement system should have a good experience in some sort of purchasing or procurement role; also, he/she should have good communication and negotiation skills. The operator must have a fundamental knowledge of sales and how marketing works, and strong persuasive skills on using automated bidding systems [6]. An experience in business and finance is also required by the operator in order to be a qualified procurement operator. The abilities to manage a budget and grow a business will be valued in a company that is seeking someone to fill this role. Once a person is qualified to become a procurement specialist, and acquire adequate experience with procurement systems, then the operator shown understand most activities in the procurement process such as manage spending, provide pricing leadership, and deal with procurement contracts. People in this job should be well educated and have strong knowledge on using E-procurement systems as well as the principles of procurement of goods and service in public bidding. Many procurement operators have a targeted area in which they work. Contracting officers promote life-cycle management for an organization and provide new, innovative services to clients. Policy analysts advise departments and officials on how to approach policies. Material management specialists organize the materials and plan the events for government programs. Real property specialists manage the life cycle of property and assets of a company or group. 1.2 User’s Perception In much of the E-procurement literature, the reduction in bidding workflow become more important than ever due to the increased complication of procurement of goods and strong competition [7, 8]. Whilst, high levels of perception to bid online using E-procurement system may be aided by increased levels of transparency [9, 10], where perceived E-procurement is low, and users may find ways to
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia
1477
avoid official purchase processes [11]. However, to date, the evidence for a positive relationship between perception of user and E-procurement compliance remains largely subjective. E-procurement offers huge potential for lowering prices paid for goods and services, and the costs associated with the purchasing process. Whilst the arrival of E-procurement creates significant potential for reducing purchasing costs, the realization of these savings requires a strong experience of users towards systems and contracts. In turn, levels of compliance appear to be influenced by user perceptions of E-procurement. However, much of this potential remains unfulfilled due to the failure of end-users to comply with systems or contracts [11]. User’s perception is hard to force. Increasingly, academics are positing the view that user perceptions of E-procurement appear to play a significant role in influencing levels of compliance with the whole procurement and bidding process [12]. However, to date, the evidence for such claims remains largely anecdotal. Preliminary research on E-procurement [13] also revealed that perceptions regarding the benefits, costs and risks of E-procurement systems significantly affect the adoption of E-procurement.
system [5]. This means that the majority of suppliers (E-procurement users) is not familiar with using E-procurement due to inefficient procurement using Internet technology, also, this result show a weak perception and awareness to bid online for public projects. In addition to that, the difficulty and expected risks involved in E-procurement activities are frequently misunderstood as a reason of failure by suppliers (E-procurement users) [4]. In this situation both experience of E-procurement operators (buyers) and the perception of users (suppliers) is fundamental to achieve high efficiency in electronic procurement process and increase the advantages of the system. This paper discusses the following two major problems: (1) Users to E-procurement systems have weak perception about online bidding and, therefore, show weak intention to participate in public bidding using E-procurement systems; (2) Few operators have adequate level of experience and background about E-procurement technology. Therefore, any attempt to adopt E-procurement in public agencies will be very difficult due to lake of expertise and qualified operators as well as reduce the efficiency and advantages of E-procurement.
2. The Problem Statement
3. The Hypothesis
The experience on using E-procurement systems from the perspective of seller and buyers is highly important to get the desired benefit from bidding and procurement of goods and services using electronic procurement systems, moreover the experience of operators on using electronic procurement systems and perception of suppliers are essential to increase the efficiency of E-procurement in public agencies. As an example, Malaysia has issued a statement calling for all 35,000 government suppliers to use its E-procurement system [5]. Nearly, all suppliers (E-procurement users) to the government have registered with E-procurement, but fewer than 18,000, electronic catalogue items have been uploaded to the
The study investigates the following two hypotheses: (1) User’s perceptions have a positive relationship with E-procurement efficiency; (2) Operator’s Experience has a positive relationship the efficiency of E-procurement process.
4. Related Literature There is a rich body of literature on E-procurement. Our review of the E-procurement literature indicates that most studies focus either on the factors affecting adoption decision of organizations for introducing E-procurement systems or the impact of E-procurement systems on organizational performance.
1478
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia
The value of E-procurement adoption and its relationship with experience of seller and buyers is defined as benefits from its implementation over costs “Perceived benefits” is a construct tied to an assessment of the gains that accrue to an individual or firm by using the technology [14]. A little attention has been given to understand the acceptance of E-procurement systems by the employees in organizations that have introduced these systems. Within the literature, a number of authors note the importance of E-procurement perception. The broad idea posited in these studies is that if perception/adoption is limited, so too are the financial benefits of E-procurement [10, 12, 15, 16]. Therefore, if the potential value of investment is to be achieved, it is critical to get potential users to adopt E-procurement systems and know how to use it when purchasing goods and services [6]. In examining influence tactics for E-procurement, attractive business cases “evaporate” when user adoption is limited [16]. They suggest that benefits are only achieved when individuals use systems and their contracts appropriately. Scholars [11] argue that E-procurement enables purchasing departments to exert increased control over organizational procurement. In line with Ref. [9], they suggest that systems increase transparency and help to point users to the approved supplier or contract. The issues of increased perception of users in ensuring appropriate use to E-procurement system [10] has been identified the term “maverick spending” incorporates the failure of individuals to use an E-procurement system when placing orders (system perception) and the failure to use mandated contracts within the system (contract compliance). 4.1 Operator Experience A procurement specialist or operator is the person within a company or organization who is responsible for acquiring the supplies, equipment, materials, property, and services necessary for the business’
existence. He or she works with the company’s financial department to secure the purchases, negotiate contracts with vendors, and implement cost-reducing methods. This person should be aware of the organization’s budget and strive to match his or her plans with the company’s financial situation. To become a procurement specialist, an individual should have a background in business, marketing or law [12]. The operator who runs E-procurement system should have a good experience in some sort of purchasing or procurement role; also, he/she should have good communication and negotiation skills. The operator must have a fundamental knowledge of sales and how marketing works, and strong persuasive skills on using automated bidding systems [6]. An experience in business and finance is also required by the operator in order to be a qualified procurement operator. The abilities to manage a budget and grow a business will be valued in a company that is seeking someone to fill this role. Once a person is qualified to become a procurement specialist, and acquire adequate experience with procurement systems, then the operator shown understand most activities in the procurement process such as manage spending, provide pricing leadership, and deal with procurement contracts. People in this job should be well educated and have strong knowledge on using E-procurement systems as well as the principles of procurement of goods and service in public bidding. Many procurement operators have a targeted area in which they work. Contracting officers promote life-cycle management for an organization and provide new, innovative services to clients. Policy analysts advise departments and officials on how to approach policies. Material management specialists organize the materials and plan the events for government programs. Real property specialists manage the life cycle of property and assets of a company or group. 4.2 User’s Perception In much of the E-procurement literature, the
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia
reduction in bidding workflow become more important than ever due to the increased complication of procurement of goods and strong competition [7]. Whilst, high levels of perception to bid online using E-procurement system may be aided by increased levels of transparency [9], where perceived E-procurement is low, users may find ways to avoid official purchase processes [11]. However, to date, the evidence for a positive relationship between perception of user and E-procurement compliance remains largely subjective. E-procurement offers huge potential for lowering prices paid for goods and services, and the costs associated with the purchasing process. Whilst, the arrival of E-procurement creates significant potential for reducing purchasing costs, the realization of these savings requires a strong experience of users towards systems and contracts. In turn, levels of compliance appear to be influenced by user perceptions of E-procurement. However, much of this potential remains unfulfilled due to the failure of end-users to comply with systems or contracts [11]. User’s perception is hard to force. Increasingly, academics are positing
the
view
that
user
perceptions
of
E-procurement appear to play a significant role in influencing levels of compliance with the whole procurement and bidding process [12]. However, to date, the evidence for such claims remains largely anecdotal preliminary research on E-procurement [13]. Also,
researcher
revealed
that
E-procurement,
perceptions, benefits, costs and risks of systems, significantly
is
affecting
the
adoption
of
E-procurement.
5. The Methodology The research applies a quantitative analysis through distributing questionnaires to 80 participants, the study population comprise users working in firms participated in online bidding for public agencies projects, and employees who are working in E-procurement department in three government
1479
administrations. Analysis of survey data provides strong empirical support for the view about the impact of perception and experience on the efficiency of E-procurement applications.
6. Results and Analysis 6.1 User’s Perception The role of user-perceived E-procurement has been discussed in E-procurement literature; with more details, the reduction in spending is taken as the main motivation for participating in E-procurement [7, 8, 12]. However, others note the apparent relationship between user-perceived E-procurement and the level of compliance, whilst, high levels of system compliance may be aided by increased levels of transparency, where Perceived E-procurement is low; users may find ways to avoid official purchase processes through E-procurement [11]. However, to date, the evidence for a positive relationship between perceived p E-procurement compliance remains largely subjective. This study assumes that “user’s perceptions has a positive relationship with E-procurement efficiency”, therefore, the study investigates this hypothesis through the following phrases: (1) The perception and attitude of users is important for swift adoption of E-procurement in public agencies: We asked the participants about their opinions with regard to the impact user’s perception to use E-procurement systems and its relationship with the adoption of E-procurement. The results shows that (3 strongly disagree, 5 disagree, 10 neutral, 19 agree, 43 strongly agree). The results statistics are shown in Fig. 1 as below: As shown from the figure, above that 95% of the participants are either agreed or strongly agree on the impact of user’s perception on E-procurement adoption. The majority of participants confirmed that adopting E-procurement is not an easy process and require that the suppliers who participate in public bidding electronically have adequate willingness and perception about how to use the system and upload
1480
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia
User’s perception about E‐procurement Number of participants
50 40 30 20 10 0 Srongly disagree
Fig. 1
Disagree
Neutral
Agree
Strongly agree
User’s perception about E-procurement.
catalogs and proposals. Whilst, the advent of E-procurement creates significant potential for reduced purchasing costs, the realization of these savings requires the commitment of internal users towards systems and contracts [17]. (2) The level of trust is low in bidding online through E-procurement. Today, IT has a major influence on commercial activities, accelerating the adoption of E-procurement and e-marketplace participation in many industries. Trust is found an effective variable between the relationship of E-procurement adoption and participation of suppliers [18]. We asked the participants about their opinion about trust to use E-procurement bidding to public projects, and measure how trust affects the user’s perception. The results shows that (0 strongly disagree, 1
disagree, 3 neutral, 21 agree, 55 strongly agree). The results statistics is shown in Fig. 2 as below: As shown in Fig. 2, there is a big increase in participants’ percentage those who agree or strongly agree with this phrase. The result shows that the vast majority of participants agrees or strongly agrees that trust is low to use E-procurement by suppliers, the mean value equal to (4.6250) and small standard deviation (0.62389) shows that public agencies should spend more efforts to convince users to use public bidding through E-procurement portals providing on the websites of public agencies. This result shows that trust on IT power of E-procurement and the power of supplier to use this technology for bidding are factors leading to adoption behavior [19]. This study suggests that trust encourage firm for more participation in electronic bidding and enhance
Number of participants
Trust to use E‐procurement 60 50 40 30 20 10 0 Srongly disagree
Disagree
Neutral
Agree
Fig. 2 Implementation of e-commerce and its impact on customer retention rate.
Strongly agree
The Impact of o Experience e and Percepttion on the Efficiency of E-procuremen E nt in Malaysia a
the perception of userss who practticing electrronic bidding throough E-procuurement porttals providedd by public agenccies. 6.2 Operator’s Experiencce o an One of the most imporrtant keys to the success of a to gain E-procuremeent system is the ability organizationn-wide adopttion by makiing it easierr for employees too order what they t need, maanagers to revview and approvee those requeests, and department headds to see the impaact of spendinng on their respective budggets. In the folloowing sectionn, the study investigates the impact of operator’s o expperience on the t efficiencyy of E-procuremeent applicatioon. This studyy assumes thaat “operator’ss experience has h a positive relaationship the efficiency of o E-procurem ment process”, therefore, t thhe study innvestigates this hypothesis thhrough the foollowing phraases: (1) The experience e off employees in i buyer’s sidde is not essentiall to ensure eff fficient electroonic procurem ment process: We askedd the particippants about thheir opinion with w regard to thhe impact exxperienced em mployees onn the efficiency off E-procurem ment system. The resullts showed that t (5 strongly disagree, 37 disagree, 366 neutral, 2 agree, a 0 stronngly agree). The results statisstics is shownn in Fig. 3 as below: b As show wn from Figg. 3 that the t majorityy of participants either does not n agree or not decided that
abo out this quesstion. The peercentage off participantss thosse who disaagree and strrongly disagrree equal too 52.5 50%, which means theree is common agreementt that the ex betw ween the participants p xperience off emp ployees is veery important to ensure effficient usagee to E-procuremen E nt. This resultt is identical with w the meann (2.4 4375) and a small standdard deviation (0.65301),, whiich increasess the reliability of our conclusion,, exp perience of em mployees is onne of the main factors thatt pub blic agenciess should loook for befo ore adoptingg E-p procurement application. a H However, insstalling latestt verssions of E-pprocurement is not enou ugh to grantt successful praccticing of online bidd ding. Highlyy exp perienced empployees who rrun the system m is essentiall and d a must requiirement for E E-procuremen nt. (2 2) There is a positive rrelationship between thee num mber of exxperienced eemployees and a numberr processing rates for bids: We W asked thee participantss about theirr perspectivee with h regard to the t number oof experienceed employeess who o operate thee E-procurem ment system, and whetherr incrreasing the number n will bbe reflected positively p onn the number of prrocessed bidss The T results showed s that (0 strongly disagree, 0 disaagree, 7 neuttral, 44 agreee, 29 strongly y agree). Thee resu ults statistics is shown in F Fig. 4 as belo ow: The T result shhows that thee majority off participantss agreees and stroongly agrees that when the numberr of experienced employees who are ablle to operate
E Employee's ex xperience and efficiency of E‐ procurementt 3% 0% 6% %
S Srongly disagree D Disagree
45% 46%
N Neutral A Agree S Strongly agree
Fig. 3
1481
The im mpact of experrienced employyees on the effficiency of E-prrocurement.
1482
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia
number of participants
Experience impact of rate of processing bids 50 45 40 35 30 25 20 15 10 5 0 Srongly disagree
Fig. 4
Disagree
Neutral
Agree
Strongly agree
The number of experienced employees on E-procurement.
E-procurement application effectively will positively reflect on the number of processed bid for final evaluation. As a result, the public agencies who frequently announce public tendering should select only qualified employees to run E-procurement system because procurement is an important and expensive business activity for organizations [2]. This is because organizations usually spend a large portion (even up to 70%) of their revenue/operational budget on purchasing goods and services [20] and only experienced employees could use a complex bidding application like E-procurement in order to avoid errors and slow processing which might be costly for both seller and buyer.
7. Conclusions E-procurement uses the Internet and the web technology in particular to replace traditional bidding and tendering processes with online ones. In order to get the maximum advantage from E-procurement applications, the public agencies who announce for
The result shows that highly experienced employees who run the system are essential and a must requirement for E-procurement. In addition to that, public agencies should spend more efforts to convince users to use public bidding through E-procurement portals providing on the websites of public agencies. This result shows that trust on IT power of E-procurement is a strong motivation to increase the perception of users of E-procurement. For an E-procurement system to be successful, it should create a true experience that allows employees to focus on their day jobs without sacrificing the visibility and management needs to effectively control spending. For employees, a highly intuitive user interface with uncluttered screens and minimal input requirements make using E-procurement system easier, even for infrequent users.
References [1] [2]
public tenders should understand that certain factors affect the whole online process including bidding, operation and using the system by the suppliers or
[3]
bidders. The experience of the employees who run E-procurement applications as well as the perceptions of the users (suppliers) are found the main determents of E-procurement success that ensure swift adoption and efficient operating the system at the same time.
[4]
P. Baily, Procurement principles and management, Prentice Hall Financial Times, England, 2008, p. 394. J. Chan, M. Lee, SMEs E-procurement adoption in Hong Kong: The roles of power, trust and value, in: Proceedings of the 36th Hawaii International Conference on Systems Science, Hawaii, USA, 2007. D.M. Gupta, R. Palmer, Moving procurement systems to the Internet: Adoption and use of E-procurement technology, European Management Journal 21 (2003) 11-23. Asian Development Bank, Inter-American Development Bank and World Bank, Strategic electronic government procurement—strategic overview: An introduction for
The Impact of Experience and Perception on the Efficiency of E-procurement in Malaysia
[5]
[6]
[7]
[8]
[9] [10]
[11]
[12]
[13]
executives, Washington, World Bank, DC, USA (2004). W. Yusoff, A. Zakaria, I. Aminul, Y.D. Hasliza, Electronic government procurement adoption behavior amongst Malaysian SMEs, International Business Research 4 (2011) 100-111. K. Arbin, The electronic procurement adoption and usage model, in: 15th Annual IPSERA Conference, San Diego, USA, April, 6-8, 2006. L. Boer, J. Harink, G. Heijboer, A conceptual model for assessing the impact of electronic procurement, European Journal of Purchasing and Supply Management 8 (2002) 25-33. S. Croom, A.B. Jones, Progress on E-procurement: Experiences from implementation in the UK Public Sector, Journal of Purchasing and Supply Management 13 (2007) 294-303. D. Neef, E-procurement: From Strategy to Implementation, Pretice Hall, NJ, USA, 2001. C. Subramaniam, M. Shaw, The effect of process characteristics on the value of B2B E-procurement, Information Technology and Management 5 (2004) 161-180. S. Croom, R. Johnston, E-service: Enhancing internal customer service through E-procurement, International Journal of Service Industry Management 14 (2003) 539-555. J. Harink, Internet-technology in Koop: De kinderschoenen ontgroeid, Dissertation, University Twente, Netherlands, 2003. H. Min, P. Galle, E-purchasing: Profiles of adopters and non-adopters, Industrial Marketing Management 32 (2003) 227-233.
1483
[14] N. Ndubisi, S. Richardson, Executive support systems in small and medium enterprises: The effect of system’s attributes and computing skill, in: Academy of International Business (AIB) & South East & Australia Regional Conference Proceeding, Shanghai, 2002. [15] A. Cox, D. Chicks, P. Ireland, T. Davies, Sourcing indirect spend: A survey of current internal and external strategies for non-revenue-generating goods and services, The Journal of Supply Chain Management 41 (2005) 39-51. [16] M. Reunis, E.M. Raaij, Scale development for E-procurement adoption influence tactics, in: Proceedings of the 15th IPSERA Conference, University of San Diego, 2006. [17] A.B. Jones, Improving E-procurement compliance: The role of user perceptions, in: POMS 20th Annual Conference, Orlando, 2009. [18] C. Hsin, W.K. Hong, Adoption of E-procurement and participation of e-marketplace on firm performance: Trust as a moderator, Information & Management 47 (2010) 262-270. [19] K. Joyce, Y. Chan, K. Mattew, O. Lee, SME E-procurement adoption in Hong Kong—The roles of power, trust and value, in: Proceedings of the 36th Hawaii International Conference on System Sciences, Hawaii, USA, Jan. 6-9, 2003. [20] J. Gebauer, A. Segev, Assessing Internet based procurement to support the virtual enterprise, in: Proceedings of the CALS Expo International and 21st century Commerce: Global Business Solutions for the New Millennium, Long beach, CA, USA, 2010.
Journal of Communication and Computer 10 (2013) 1484-1496
Optimal Object Categorization under Application Specific Conditions Steven Puttemans1, 2 and Toon Goedeme1, 2 1. Embedded and Applied VISion Engineering Research Group, Faculty of Engineering Technology, University of Leuven Campus De Nayer, Sint Katelijne Waver 2860, Belgium 2. ESAT/PSI-VISICS Research Group, Faculty of Engineering, University of Leuven, Heverlee 3001, Belgium
Received: October 14, 2013 / Accepted: November 15, 2013 / Published: December 31, 2013. Abstract: Day-to-day industrial computer vision applications focusing on object detection have the need of robust, fast and accurate object detection techniques. However, current state-of-the-art object categorization techniques only reach about 85% detection rate when performing in the wild detections, which try to cope with as much scene and object variation as possible. However, several industrial applications show many known characteristics like constant lighting, known camera position, constant background …, giving lead to several constraints on the actual algorithms. With a complete new universal object categorization framework, we want to prove the detection rate of these object categorization algorithms by exploiting the application specific knowledge which can help to reach a robust detector with detection rates of 99.9% or higher. We will use the same constraints to effectively reduce the number of false positive detections. Furthermore, we will introduce an innovative active learning system based on this application specific knowledge that will drastically reduce the amount of positive and negative training samples, leading to a shorter and more effective annotation and training phase. Key words: Universal object categorization, application specific constraints, innovative active learning.
1. Introduction Many industrial companies still use classic object detection approaches like pixel based segmentation, template based matching or using an actual object CAD model to look for an object in an image. These approaches work well when looking for specific objects but when trying to detect object candidates from a complete object class, these techniques start to fail rather soon. The main problem lies in the natural variation of the products, also called the intra-class variability. This means that within a single object class (e.g. the class of apples), an object can vary in size, shape, color …, However, more recent research on object detection has resulted in a set of Corresponding author: Steven Puttemans, M.Sc. Artificial Intelligence, research assistant and Ph.D. student, research fields: computer vision, machine learning and artificial intelligence. E-mail: steven.puttemans@kuleuven.be.
state-of-the-art object categorization techniques [1-4], which focus on creating a general model of an object class and not an object specific model for each instance. These techniques mainly focus on in the wild detections and try to cope with as much variation as possible by using variation invariant image formats (e.g. the use of histogram of oriented gradients in detection of pedestrians or cars). Many industrial applications however, put some sort of restriction on the application setup, like a constant lighting, a known object position, a small range of object scales, a known camera position …. Using this knowledge of application specific scene and object variation, this research aims at creating a universal object categorization framework which can reach detection rates up to 99.9%. This very high detection rate is one of the many requirements of industrial applications, before the industry will even
Optimal Object Categorization under Application Specific Conditions
consider using object categorization techniques. The Ph.D. research will try to incorporate these known restrictions into an actual innovative object categorization framework. This paper is organized as follows: Section 2 discusses the research problem in more detail; Section 3 discusses the state-of-the-art of object categorization techniques; Section 4 describes the objectives of this PhD research; Section 5 contains the proposed methodology; Section 6 finally contains a conclusion and discusses the expected outcome.
2. Research Program The focus of this research lays in industrial computer vision applications that want to perform object detection on object classes with a high intra-class variability. This means that objects have varying size, color, texture, orientation ..., Examples of these specific industrial cases can be seen in Fig. 1. These day-to-day industrial applications, such as product inspection, counting and robot picking, are in desperate need of robust, fast and accurate object detection techniques which reach detection rates of 99.9% or higher. However, current state-of-the-art object categorization techniques only guarantee a detection rate of Âą85% when performing in the wild detections [5]. In order to reach a higher detection rate, the algorithms impose very strict restrictions on the actual application environment, e.g. a constant and uniform lighting source, a large contrast between objects and background, a constant object size and color ..., Compared to these more complex object
1485
categorization algorithms, classic thresholding based segmentation technique require all of these restrictions to even guarantee a good detection result and are unable to cope with variation in the input data. Looking at the state-of-the-art object categorization techniques, we see that the evolution of these techniques is driven by in the wild object detection (Table 1). The main goal exists in coping with as many variations as possible, achieving a high detection rate in very complex scenery. However, specific industrial applications easily introduce many constraints, due to the application specific setup of the scenery and the objects. Exploiting that knowledge can lead to smarter and better object categorization techniques. For example, when detecting apples on a transportation system, many parameters like the location, background and camera position are known. Current object categorization techniques do not use this information because they do not expect this kind
Fig. 1 Examples of industrial object categorization applications: Robot picking and object counting of natural products (checking flower quality, picking pancakes, counting micro-organisms, picking peppers).
Table 1 Evolution in robustness of object recognition and object detection techniques trying to cope with object and scene variation as mentioned in Ref. [6].
(1. Illumination differences; 2. Location of objects; 3. Scale changes; 4. Orientation of objects; 5. Occlusions; 6. Clutter in scene; 7. Intra-class variability), referenced papers [2, 6-9].
1486
Optimal Object Categorization under Application Specific Conditions
of known variation. However, exploiting this information will lead to a new universal object detection framework that yields high and accurate detection rates, based on the scenery specific knowledge.
3. State-of-the-Art Object detection is a widely spread research topic, with large interest in current state-of-the-art object categorization techniques. Ref. [1] suggested a framework based on integral channel features, where all object characteristics are captured into feature descriptions which are then used as a large pool of training data in a boosting process [10]. In contrast to the original boosted cascade of weak classifiers approach, suggested by Ref. [4], this technique incorporates multiple sources of information to guarantee a higher detection rate and less false positive detections. In the following years of research, this technique has been a backbone for many well performing object detection techniques, mainly for into the wild detections of pedestrians [5, 11, 12] and traffic signs [13]. All these recently developed techniques profit from the fact that the integral channel features framework allows integrating extra application-specific knowledge like stereo vision information, knowledge of camera position, ground plane assumption ..., to obtain higher detection rates. The concept of using application specific scene constraints to improve these state-of-the-art object categorization techniques was introduced in Ref. [14]. The paper suggests using the knowledge of the application specific scene and object conditions as constraints to improve the detection rate, to remove false positive detections and to drastically reduce the number of manual annotations needed for the training of an effective object model. Aside from effectively using the scene and object variation information to create a more accurate application specific object detector, the Ph.D. research
will focus on reducing the amount of time needed for manually annotating gigantic databases of positive and negative training images. This will be done using the technique of active learning, on which a lot of recent research was performed [15, 16]. This research clearly shows that integrating multiple sources of information into an active learning strategy can help to isolate the large problem of outliers giving reason to include the wrong examples.
4. Outline of Objectives During this Ph.D. existing state-of-the-art object, categorization algorithms will be reshaped into a single universal semi-automatic object categorization framework for industrial object detection, which exploits the knowledge of application specific object and scene variation to guarantee high detection rates. Exploiting this knowledge will enable three objectives, each focusing on another aspect of object detection that is important for the industry. (1) A high detection rate of 99.9% or even higher: Classic techniques reach detection rates of 85% during in the wild detections, but for industrial applications a rate of 99.9% and higher are required. By integrating the knowledge of the object and scene variation, the suggested approach will manage to reach this high demands. Using the framework of Ref. [1] as a backbone for the universal object, categorization framework will be created, and these characteristics will be used to include new feature channels to the model training process, focusing on this specific object and scene variation. (2) A minimal manual input during the training of an object model: Classic techniques demand many thousands of manual annotations during the collection of training data. By using an innovative active learning strategy, which again uses the knowledge of application specific scene and object variation, the number of manual annotations will be reduced to a much smaller number of input images. By iteratively annotating only a small part of the training set and
Optimal Object Categorization under Application Specific Conditions
using that to train a temporary detector based on the already annotated images, the algorithm will decide which new examples will actually lead to a higher detection rate, only offer those for a new annotation phase and omit the others. (3) A faster and more optimized algorithm: By adding all of this extra functionality, resulting in multiple new feature channels, into a new framework, a large portion of extra processing is added. Based on the fact that the original algorithm is already time consuming and computational expensive, the resulting framework will most likely be slower than current state-of-the-art techniques. However, by applying CPU and GPU optimizations wherever possible, the aim of the Ph.D. is to still provide a framework that can supply real time processing. The use of all this application specific knowledge from the scene and the object, with the aim of reaching higher detection rates, is not a new concept. Some approaches already use pre- and post- processing steps to remove false positive detections based on application specific knowledge that can be gathered together with the training images. For example, Ref. [11] uses the knowledge of a stereo vision setup and ground plane assumption, to reduce the area where pedestrian candidates are looked for. This Ph.D. research, however, will take it one step further and will try to integrate all this knowledge into the actual object categorization framework. This leads to several advantages over the pre- and postprocessing approaches: There will be no need for manual defining or capturing features, which are interesting for these preand postprocessing steps. Multiple features will be supplied as a large package to the framework. The underlying boosting algorithm will then decide which features are actually interesting to use for model training. The algorithm can separate the input data better than human perception based on combination of features. Each possible scene and object variation will be
1487
transformed into a new feature channel, in order to capture as much variation as possible. Once a channel is defined, it can be automatically recalculated for every possible application. Besides not being able to reach top level detection rates, state-of-the-art object categorization techniques face the existence of false positive detections. These detections are classified by the object detector model as actual objects, because they contain enough discriminating features. However, they are no actual objects in the supplied data. By adding a larger set of feature channels to the framework, and integrating a larger knowledge of scene and object variation during the training phase, the resulting framework will effectively reduce the amount of false positive detections.
5. Methodology In order to ensure a systematic approach, the overall research problem of the Ph.D. is divided into a set of subproblems, which can be solved one by one in an order of gradual increase in complexity, in order to guarantee the best results possible. Section 5.1 will discuss the integration of the application specific scene and object variation during the model training process, by highlighting different variation aspects of possible applications and how they will be integrated as feature channels. Section 5.2 will illustrate how the use of an innovative active learning strategy can help out with reducing the time consuming job of manual annotation. Finally, Section 5.3 will discuss how the resulting framework can be optimized using CPU and GPU optimizations wherever possible. 5.1. Integration of Scene and Object Variation during Model Training Different properties of application specific scene and object variation allow designing a batch of new feature channels in a smart way, that can be used for a universal object categorization approach. During training the generation is stimulated as many extra
1488
Optimal Object Categorization under Application Specific Conditions
feature channels (Fig. 2) as possible, in order to capture as many variation and knowledge of the application as possible from the image data. This is no problem, since the boosting algorithm of the training will use all these features to determine which feature channels capture the most variation, in order to prune channels away and only keep the most descriptive feature channels. This immediately ensures that the algorithm will not become extremely slow during the actual detection phase because of the feature channel generation. By integrating all these extra feature channels into the actual object model training process, a better universal and more accurate object categorization framework will be supplied, which works very application specific to reach the highest performance and detection rate possible. In Subsection 5.1.1, the influence of the object scale and position in the image will be discussed. Subsection 5.1.2 discusses the influence of lighting, color and texture. Subsection 5.1.3 addresses the influence of background clutter and occlusion. Finally Subsection 5.1.4 will handle the object rotation and orientation knowledge. 5.1.1. Influence of Object Scale and Position In state-of-the-art object categorization, an object model is trained by rescaling all provided training images towards a fixed scale, which results into a single fixed scale model. Using a sliding window approach, with the window size equal to the size of the resulting model, object detection is performed at each image position. However, there are only a limited number of applications that have fixed scale objects. In order to detect objects of different scales in all those other applications, an image scale space pyramid is generated. In this scale, space pyramid the original image is down and up sampled and used with the single scale model. This will generate the possibility to detect objects at different scales, depending on the amount of scales that are tested. The larger the pyramid, the more scales that will be tested but the longer the actual detection phase will take.
Reducing this scale, space pyramid effectively is a hot research topic. Ref. [5] interpolates between several predefined images scales, while the detector of Ref. [11] uses an approach that interpolates between different trained scales of the object model. These multiscale approaches are frequently used because the exact range of object scales is unknown beforehand in many applications. However, many industrial applications have the advantage that the position of the complete camera setup is fixed and known beforehand (e.g. a camera mounted above a conveyor belt). Taking this knowledge into account, the scale and position of the objects can actually be computed and described fairly easy as seen in Fig. 3). Using this information, new feature channels can be created. Based on manual annotation information, a 2D probability distribution can be produced over the image giving a relation between the scale and the position of the object in the image. Ref. [17] discusses a warping window technique that uses a lookup function defining a fixed rotation and a fixed scale for each position in the
Fig. 2 Example of different image channels used in the integral channel features approach of Ref. [1]. (a) Grayscale image; (b) LUV color space; (c) Gabor orientation filters; (d) Difference of Gaussians; (e) Gradient magnitude; (f) Edge detector; (g) Gradient histogram; (h) Thresholded image.
Optimal Object Categorization under Application Specific Conditions
1489
Fig. 3 Left: Example of a scale-location-rotation lookup function for pedestrians in a fixed and lens deformed camera setup; Right: Example of a fragmented scale space pyramid.
image. However, reducing the detection to a single scale for each position limits the intra-class variability that object categorization wants to maintain. To be sure this is not a problem, instead of using a fixed scale, a probability distribution of possible scales for each position can be modeled. The use of these distribution functions can lead to a serious reduction of the scale space pyramid, resulting in a fragmented scale space pyramid, as seen in Fig. 3. This fragmented scale space pyramid can again be used as a separate feature channel for object model training. 5.1.2. Influence of Lighting, Color and Texture State-of-the-art object categorization ensures a certain robustness by making training samples and new input images invariant for color and lighting variations. To do so, they use a color invariant image form, like a histogram of oriented gradient representation. Another possible approach is to use Haar-like features, like suggested by Ref. [4]. Making the images invariant to lighting and color has a twofold reason. First of all, the color variation in academic application is too large (e.g. the colors of clothing in pedestrian detection). Secondly, the color is too much influenced by the variation in lighting conditions. Therefore, academic applications try to remove as many of this variation as possible by applying techniques like histogram equalization and the use of gradient images. By choosing a color and light invariant image form, all the information from the RGB color spectrum is lost which is in fact quite useful in the industrial applications suggested by this Ph.D. research. In many
of these applications, a uniform and constant lighting is used, leading to fixed color values. This information cannot be simply ignored when detecting objects with specific color properties like strawberries. The advantage of adding this color information has already been proven in Ref. [1], where color information of the HSV and LUV space is added to obtain a better and more robust pedestrian detector. Besides focusing on the color information, it can be interesting to focus on multispectral color data. It is possible that objects cannot be separated in the visual RGB color spectrum, but there are higher multispectral frequency resolutions that make the separations of objects and background rather easy. Academic research [18-20] has already shown great interest in these multispectral approaches, where most of the applications are located in remote sensing and mobile mapping. Another parameter that is not widely spread for object categorization is the use of relevant texture information in the training objects. Texture can be described as a unique returning pattern of gradients, which will almost never occur in the background information. In order to derive these patterns from the input data, techniques like Fourier transformations [21] (Fig. 4) and Gabor filters [22] are used. These transformations show which frequencies are periodically returning in the image to define application and object specific textures. 5.1.3. Influence of Background Clutter and Occlusion State-of-the-art object categorization approaches
1490
Optimal Object Categorization under Application Specific Conditions
Fig. 4 Texture variation based on the Fourier power spectrum of an orange and a strawberry.
always attempt to detect objects in the wild, which means that it can occur in every kind of situation, leading to an infinite number of possible background types ..., In order to build a detector that is robust to all this scene background variation, an enormous amount of negative images samples is needed during model training. This is required to try to model the background variation for correct classification and to ensure that the actual object model will not train background information. Besides that, it is necessary to collect as much positive examples as possible in those varying environments. Doing so ensures that only object features get selected that describe the object unrelated to the background behind it. This variation in the background is referred to as clutter. Many industrial applications, however, have a known background, or at least a background with minimal variation. Combined with occlusion, where the object is partially or completely covered, clutter seems to happen much less frequent than in the wild detection tasks. Take for example the taco’s on the conveyor belt in Fig. 5. The conveyor belt is moving and changes maybe slightly, but it stays quite constant during processing. Making a good model of that background information, can help to form an extra feature channel defining foreground and background information. Other cases, like the picking of pears, will have much more
variation in background, and will not give the possibility to simply apply foreground-background segmentation (Fig. 6). A technique that is widely used for this kind of information is foreground-background segmentation, like in Ref. [23]. This technique helps us identify regions in the image that can be classified as foreground and thus regions of interest for possible object detections. The masks created by this segmentation can be applied as an extra feature channel. Using a dynamic adapting background model [24], the application specific background will be modeled and a likelihood map of a region belonging to the foreground will be created. These are referred to as heat maps.
Fig. 5 Example of background variation and occlusion in (a) academic cases and (b) industrial cases.
Fig. 6 Example of pear fruit in an orchard, where more background clutter and occlusion occurs.
Optimal Object Categorization under Application Specific Conditions
Due to the context of application specific algorithms, one can state that the only negative images that need to be used as negative training samples are images that contain the possible backgrounds. This leads to the conclusion that many case specific object models can be reduced to having a very limited amount of negative training images, based on the applications scene and background variation, maybe even reducing the negative training images to a single image, if a static background occurs. 5.1.4. Influence of Rotation and Orientation Most state-of-the-art object categorization approaches, e.g. detecting pedestrians, assume that there is no rotation of the actual object, since pedestrians always appear more or less upright. However, this is not always the case, like shown in Ref. [17], where pedestrians occur in other orientations due to the lens deformation and the birdseye viewpoint of the camera input (Fig. 7). Many industrial applications, however, contain different object orientations, which leads to problems when having a fixed orientation object model. Adding all possible orientations to the actual training data for a single model, will lead to a model that is less descriptive and which will generate tons of extra false positive detections. A second approach is to test all possible orientations, by taking a fixed angle step, rotating the input image and then trying the trained single orientation model. Once a detection is found, it can be coupled to the correct angle and then used to rotate the detection bounding box, like discussed in Ref. [25]. However, in order to reach real-time performance using this approach, a lot of GPU optimizations will be needed, since the process of
1491
rotating and performing a detection on each patch is computationally intensive. A possible third approach trains a model for each orientation, as suggested in Ref. [26]. However, this will lead to an increase of false positive detections. The currently used approaches to cope with different orientations do not seem to be the best approaches possible. In this Ph.D. research, we want to create an automated orientation normalization step, where each patch is first put through a series of orientation filters that determine the orientation of the current patch and then rotates this patch towards a standard model orientation. A possible approach is the dominant gradient approach as illustrated in Fig. 8. However, preliminary test results have shown that this approach does not work in every case. Therefore, a combination of multiple orientation defining techniques will be suggested in our framework. Other techniques that can be included into this approach are eigenvalues of the covariance matrix [27], calculating the geometric moments of a color channel of the image [28] or even defining the primary axis of an
Fig. 7 Example of viewpoint and lens deformation, changing the natural orientation of objects [17].
Fig. 8 Example of rotation normalization using a dominant gradient technique. From left to right: original image (road marking), gradient image, dominant orientation and rotation-corrected image.
1492
Optimal Object Categorization under Application Specific Conditions
ellipse fitted to foreground-background segmentation data [29]. Our suggested orientation normalization filter will use the combination of multiple orientation features to decide which one is the best candidate to actually define the patch orientation. In order to create this extra filter, all manual positive annotations are given an extra parameter, which is the object orientation of the training sample. From that data a mapping function is learned to define a pre-filter that can output a general orientation for any given window. Part of this general idea, where the definition of the orientation is separated from the actual detection phase, is suggested in Ref. [30]. 5.2. Innovative Active Learning Strategy for Minimal Manual Input Limited scene and object variation can be used to put restrictions on the detector, by supplying extra feature channels to the algorithm framework, as previously explained. However, we will take it one step further. The same information will be used to optimize the complete training process and to drastically reduce the actual amount of training data that is needed for a robust detector. For state-of-the-art object categorization algorithms, the most important way to obtain a detector with a high detection rate is increasing the amount of positive and negative training samples enormously. The idea behind it is simple, if you add a lot of extra images, you are bound to have those specific examples that lie close to the decision boundary and that are actually needed to make an even better detector. However, since several industrial applications have a smaller range of variation, it should be possible to create an active learning strategy based on this limited scene and object variation, which succeeds in getting a high detection rate with as less examples as possible, by using the variation knowledge to look for those specific examples close to the decision boundary.
Like described in the conclusion of Ref. [14], using immense numbers of training samples is currently the only way to reaching the highest possible detection rates. Since all these images need to be manually annotated, which is very time consuming job, this extra training data is a large extra cost for industrial applications. Knowing that the industry wants to focus more and more on flexible automatization of several processes, this extra effort to reach high detection rates is a large downside to current object categorization techniques, since companies do not have the time to invest all this manual annotation work. The industry wants to retrieve a robust object model as fast as possible, in order to start using the detector in the actual detection process. 5.2.1. Quantization of Existing Scene and Object Variation In order to guarantee that the suggested active learning approach will work, it is necessary to have a good quantization of the actual variation in object and scene. These measurements are needed to define if new samples are interesting enough to add as extra training data. The main focus is to define how much intra-class variation there is, compared to the amount of variation in the background. Many of these variations, like scale, position, color ..., can be expressed by using a simple 1D probability distribution over all different training samples. However, some variations are a lot harder to quantize correctly. If it is important to guarantee the intra-class variability, then it can even be extended to a 2D probability distribution, to allow multiple values for a single point in the distribution. However, features like texture and background variation cannot be modeled with a simple 1D probability distribution. A main part of the Ph.D. research will go into investigating this specific problem and trying to come up with good quantizations for these entire scene and object variations. 5.2.2. Active Learning during Object Model Training Initial tests have shown that it is possible to build
Optimal Object Categorization under Application Specific Conditions
robust object detectors by using only a very limited set of data, as long as the training data is chosen based on application specific knowledge. However, figuring out which examples are actually needed, sometimes turns out to be more time consuming than just simply labeling large batches of training data, if the process is not automated. Therefore, we suggest using an active learning strategy which should make the actual training phase simpler and more interactive. Eventually the algorithm optimizes two aspects: first being a minimal manual intervention and secondly an as high as possible detection rate. This research will be the first of its kind to integrate the object and scene variation into the actual active learning process, combining many sources of scene and object specific knowledge to select new samples, which can then be annotated in a smart and interactive way. Fig. 9 shows how the suggested active learning strategy based on application specific scene and object variation should look like. As a start, a limited set of training data should be selected from a large database of unlabeled images. Since capturing many input
1493
images is not the problem in most cases, the largest problem lies in annoting the complete set, which is very time consuming. Once this initial set of data is selected, they are given to the user for annotation and a temporarily object model is trained using this limited set of samples. After the training a set of test images is smartly selected from the database using the scene and object variations that are available. By counting the true positives, false positives, true negatives and false negatives, the detector performance is validated on this test data, by manually supervising the output of the initial detector. Based on this output and the knowledge of the variation distributions in the current images, an extra set of training images is selected cleverly. The pure manual annotation is now split into a part where the operator needs to annotate a small set of images, but after the detection step, needs to validate the detections in order to compute the correctness of the detection output. This process is iteratively repeated until the desired detection rate is reached and a final object model is trained.
Fig. 9 Workflow of the suggested active learning strategy. Hand symbol = manual input, star symbol = knowledge of scene and object variation is used, TP = true positive detection, TN = true negative detection, FP = false positive detection, FN = false negative detection).
1494
Optimal Object Categorization under Application Specific Conditions
The above described innovative active learning strategy will yield the possibility to make a well fundamented guess on how many positive and negative training samples there will actually be needed to reach a predefined detection rate. In doing this, the approach will drastically reduce the amount of manual annotations that need to be provided, since it will only propose to annotate new samples that actually improve the detector. Training images that describe frequently occurring situations are classified as objects with a high certainty, which are not interesting in this case. On the contrary, it will be more interesting trying to select those positive and negative training samples that lie very close to the decision boundary, in order to make sure that the boundary will be more stable, more supported by good examples and thus leading to higher detection rates. It is important to mention that classic active learning strategies are often quite sensitive to outliers [31] that get selected in the learning process and that lead to overfitting of the training data. However, by adding multiple sources of information, being different application specific scene and object variations, the problem of single outliers can be countered, since their influence on the overall data distribution will be minimal. The suggested approach will filter out these outliers quite effectively, making sure that the resulting detector model will not overfit to the actual training set. 5.3 CPU and GPU Optimization towards a Real-Time Object Categorization Algorithm Once the universal object categorization framework, combined with an innovative active learning strategy, will be finished and it will produce a better and more accurate detection system for industrial applications and in general, for all applications where the variation in scene and/or object is somehow limited. However, expanding a framework to cope with all these application specific scene and object variations will
lead to more internal functionality. This will result in a computationally more expensive and a slower running algorithm. Since real time processing is essential for most industrial applications, this problem cannot be simply ignored. The longer the training of a specific object model takes, the more time a company invests in configuration and not in the actual detection process that generates a cash flow. This is why during this Ph.D. research each step of the processing will be optimized using CPU and GPU optimization. Classical approaches like parallelization and the use of multicore CPU’s can improve the process [32], while the influence of GPGPU (general purpose graphical processing units) will also be investigated. The CUDA language will be used to implement these GPU optimizations, but the possibility of using OpenCL will be considered
6. Conclusions and Expected Outcome At the end of this Ph.D. research, a complete new innovative object categorization framework will be available that uses industrial application specific object and scene constraints, in order to obtain an accurate and high detection rate of 99.9% or higher. The result will be a stimulation for the industry to actively use this technology for robust object detection. Industrial companies will be able to apply robust object detection in many varying applications, using the constraints of each specific application in a smart way. This research will also lead to new insights in general for object detection techniques meaning that if this approach is proved to be successful, the same approach will be introduced in other frameworks like the deformable parts model of Ref. [11], to reach higher performances without increasing the number of training examples.
Acknowledgements This work is supported by the Institute for the Promotion of Innovation through Science and
Optimal Object Categorization under Application Specific Conditions
Technology in Flanders (IWT) via the IWT-TETRA project TOBCAT: Industrial Applications of Object Categorization Techniques.
References [1] [2]
[3]
[4]
[5]
[6] [7]
[8]
[9]
[10]
[11]
[12]
[13]
P. Dollar, Z. Tu, P. Perona, S. Belongie, Integral channel features, in: BMVC, London, Sep. 7-10, 2009. P. Felzenszwalb, R. Girshick, D. McAllester, Cascade object detection with deformable part models, in: The 23th IEEE Conference on Computer Vision and Pattern Recognition, CA, Jun. 13-18, 2010. J. Gall, V. Lempitsky, Class-specific hough forests for object detection in: A. Criminisi, J. Shotton (Eds.), Decision Forests for Computer Vision and Medical Image Analysis, Springer, 2013, pp. 143-57. P. Viola, M. Jones, Rapid object detection using a boosted cascade of simple features, in: 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2001), Kauai, HI, USA, Dec. 8-14, 2001. P. Dollár, S. Belongie, P. Perona, The fastest pedestrian detector in the west, in: Proceedings of the British Machine Vision Conference 2010, UK, Agu. 31-Sep. 03, 2010. J. Lewis, Fast normalized cross-correlation, Vision Interface 10 (1995) 120-123. J. Hsieh, H. Liao, K. Fan, M. Ko, Y. Hung, Image registration using a new edge-based approach, Computer Vision and Image Understanding 67 (1997) 112-130. F. Mindru, T. Tuytelaars, L. Gool, T. Moons, Moment invariants for recognition under changing viewpoint and illumination, Moment Invariants for Recognition Under Changing Viewpoint And Illumination 94 (2004) 3-27. H. Bay, T. Tuytelaars, L. Van Gool, Surf: Speeded up robust features, in: 9th European Conference on Computer Vision, Graz, Austria, May. 7-13, 2006. Y. Freund, R. Schapire, N. Abe, A short introduction to boosting, Japanese Society For Artificial Intelligence 14 (1999) 771-780. R. Benenson, M. Mathias, R. Timofte, L. Van Gool, Pedestrian detection at 100 frames per second, in: 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012, pp. 2903-2910. R. Benenson, M. Mathias, T. Tuytelaars, L. Van Gool, Seeking the strongest rigid detector, in: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Oregon, USA, 2013. M. Mathias, R. Timofte, R. Benenson, L. Gool, Traffic sign recognition—How far are we from the solution?, in: Proceedings of IEEE International Joint Conference on Neural Networks, Oregon, USA, 2013.
1495
[14] S. Puttemans, T. Goedeme, How to exploit scene constraints to improve object categorization algorithms for industrial applications?, in: Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP 2013), Barcelona, Spain, Feb. 21-24, 2013, pp. 827-830. [15] A. Kapoor, K. Grauman, R. Urtasun, T. Darrell, Active learning with gaussian processes for object categorization, in: IEEE 11th International Conference on Computer Vision, Rio de Janeiro, Brazil, Oct. 14-21, 2007. [16] X. Li, Y. Guo, Adaptive active learning for image classification, in: IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), Portland, USA, Jun. 23-28, 2013. [17] K.V. Beeck, T. Goedeme, T. Tuytelaars, A warping window approach to real-time vision-based pedestrian detection in a truck’s blind spot zone, in: 9th International Conference on Informatics in Control, Automation and Robotics, Rome, Italy, Jul. 28-31, 2012, pp. 561-568. [18] C.O. Conaire, N.E. O’Connor, E. Cooke, A.F. Smeaton, Multispectral object segmentation and retrieval in surveillance video, in: 2006 IEEE International Conference on Image Processing, Atlanta, USA, Oct. 8-11, 2006, pp. 2381-2384. [19] A.K. Shackelford, C.H. Davis, A combined fuzzy pixel-based and object-based approach for classification of high-resolution multispectral data over urban areas, IEEE Transactions on Geoscience and Remote Sensing 41 (2003) 2354-2363. [20] Q. Yu, P. Gong, N. Clinton, G. Biging, M. Kelly, D. Schirokauer, Object-based detailed vegetation classification with airborne high spatial resolution remote sensing imagery, Journal of The American Society for Photogrammetry and Remote Sensing 72 (2006) 799-811. [21] R. Cant, C.S. Langensiepen, D. Rhodes, Fourier texture filtering, in: 2013 UKSim 15th International Conference on Computer Modelling and Simulation (UKSim), Cambridge, UK, Apr. 10-12, 2013, pp. 123-128. [22] F. Riaz, A. Hassan, S. Rehman, U. Qamar, Texture classification using rotation-and scale-invariant gabor texture features, IEEE Signal Processing Letters 20 (2013) 607-610. [23] C.H. Yeh, C.Y. Lin, K. Muchtar, L.W. Kang, Real-time background modeling based on a multi-level texture description, Information Sciences 269 (2013) 106-127. [24] M. Hammami, S.K. Jarraya, H. Ben-Abdallah, On line background modeling for moving object segmentation in dynamic scenes, Multimedia Tools and Applications 63 (2013) 899-926. [25] A. Mittal, A. Zisserman, P. Torr, Hand detection using multiple proposals, in: BMVC 2011: The 22nd British Machine Vision Conference, University of Dundee, Aug.
1496
Optimal Object Categorization under Application Specific Conditions
29-Sep. 02, 2011. [26] C. Huang, H. Ai, Y. Li, S. Lao, Vector boosting for rotation invariant multi-view face detection, in: Tenth IEEE International Conference on Computer Vision ICCV 2005, Oct. 17-21, 2005, pp. 446-453. [27] G. Kurz, I. Gilitschenski, S. Julier, U.D. Hanebeck, Recursive estimation of orientation based on the bingham distribution, arXiv preprint arXiv:1304.8019, 2013. [28] G.A. Leiva-Valenzuela, J.M. Aguilera, Automatic detection of orientation and diseases in blueberries using image analysis to improve their postharvest storage quality, Food Control 33 (2013) 166-173. [29] M.G. Ascenzi, Determining orientation of cilia in
connective tissue, US Patent, 8345946 (2013). [30] M. Villamizar, F. Moreno-Noguer, J. Andrade-Cetto, A. Sanfeliu, Efficient rotation invariant object detection using boosted random ferns, in: 2010 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), CA, Jun. 13-18, 2010, pp.1038-1045. [31] C.C. Aggarwal, Supervised Outlier Detection, in Outlier Analysis, Springer, pp. 169-198, 2013. [32] F. De Smedt, K. Van Beeck, T. Tuytelaars, T. GoedemĂŠ, Pedestrian detection at warp speed: Exceeding 500 detections per second, in: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Portland, USA, Jun. 23-28, 2013, pp. 622-628.
Journal of Communication and Computer 10 (2013) 1497-1501
Image Super Resolution and Enhancement Using E-spline Gamal Fahmy Department of Electrical Engineering, College of Engineering, University of Majmaah, Majmaah 11952, KSA Received: October 24, 2013 / Accepted: November 24, 2013 / Published: December 31, 2013. Abstract: E-splines (Exponential spline) polynomials represent the best smooth transition between continuous and discrete domains. As they are constructed from convolution of exponential segments, there are many degrees of freedom to optimally choose the most convenient E-spline, suitable for a specific application. In this paper, the parameters of these E-splines were optimally chosen, to enhance the performance of image zooming and interpolation schemes. The proposed technique is based on minimizing the total variation function of the detail coefficients of the E-spline based wavelet decomposition. In zooming applications, the quality of interpolated images are further improved and sharpened by applying ICA technique to them, in order to remove any dependency. Illustrative examples are given to verify image enhancement of the proposed E-spline scheme, when compared with the existing approaches. Key words: Image de-noising, interpolators, E-spline functions.
1. Introductionď€ During the past decade, there have been an increasing number of papers devoted to the use of polynomial splines in different signal processing applications [1-3]. B-spline polynomials, is a class of these polynomial splines that find extensive applications in many engineering applications. In Ref. [4], a complete analysis for a B-spline PR (perfect reconstruction) frame work with a derivation for the scaling and wavelet functions was presented. However, as they are constructed using Haar functions, there is no much degree of freedom to use in optimizing the performance of some signal processing applications like the design of digital interpolators. On the other hand, E-splines (Exponential splines) enjoy a unique feature of being able to convert from analog to digital applications. This is crucial in several signal processing applications such as differential operators, fractional Corresponding author: Gamal Fahmy, Ph.D., research fields: image enhancement, image de-noising and image interpolation. E-mail: fahmygamal@hotmail.com.
delays, interpolators and sampling rate converters [5-6]. Moreover, E-splines have many degrees of freedom if they are optimized in a specific application, as they are constructed from the convolution of exponential segments with different rates. In Ref. [7], a preliminary application for the usage of E-splines in image zooming and interpolation was presented. In this paper, it is proposed to use E-splines in enhancing the performance of image de-noising as well as image zooming schemes. In denoising applications, the proposed denoising technique is based on total variation function minimization [8-9]. Using a recently developed E-spline wavelet decomposition [10-14], the E-spline parameters as well as the thresholding levels of the E-spline detail coefficients are optimally chosen to minimize the total variation of the E-spline detail wavelet coefficients. In image zooming applications, E-spline based interpolators are used in image interpolation. In this case, the parameters of E-spline polynomials are chosen to boost the high frequency detail energy of
1498
Image Super Resolution and Enhancement Using E-spline
the interpolated image. Further improvement is possible by estimating the missing high frequency details from the given low resolution image using the information contained in the whole low frequency image as described in Section 4 [15-18]. Further image enhancement is also possible, by applying ICA techniques [19]; to the interpolated images in order to boost high frequency details and reduce any dependency between them. Illustrative examples are given to verify the ability of E-spline polynomials to significantly enhance image quality. The paper is organized as follows: in Section 2, a brief description of E-spline polynomials and wavelet PR systems is given. In Section 3 describes the design of a PR system using E-spline polynomials. Section 4 introduces our E-spline based super resolution methodology. Section 5 concludes the paper.
m m Bm 2 Bm 2 1 B m 1 B m m m 2 2 B 0 0
Bm (t ) B11 (t ) B1 2 (t ) ... B1 m (t )
(1)
where, B1 (t ) e , 0 t 1 .The vector can assume any positive, negative or even complex conjugate values. This means a considerable flexibility over cardinal B-spline polynomials that only use Haar functions. Bm (t ) is of finite support and equals zeros
t
at t ≤ 0 and t ≥ m. Between the knots t = 1, 2, … m – 1, it is represented by polynomials of order (m – 1) in t [5]. Due to its continuity and smoothness, it is used to expand continuous signals s(t). In the discrete case, s(n) can be expressed using the convolutional relation.
s ( n) c ( k ) Bm ( n k )
(2)
k
The ck coefficients are obtained using the concept of inverse filtering described in Ref. [4]. In Ref. [7], an alternate approach is given to determine these coefficients as the solution of the linear system
0
m ... Bm 1 2
t L
E-spline 2-scale relation defined by m
p ( k ) B ( 2t k ) , m
k m
s ( 2) ... s ( N ) t ,
c c (1) c ( 2) ... c ( N )
t
(4)
j 1 m P ( z ) p (k ) z k ` , z e 2 2 k m has been determined for any arbitrary n . Next, the wavelet E-spline function m (t ) , satisfying the orthogonality relation: m
m
( t ) B m ( t l ) dt 0
m
i .e .
m
(t )
m
q (k ) B (2t k )
km
(5)
m
1 m Q (z) q (k ) z k 2 km Moreover, the Exponential dual scaling A(z) and wavelet E-spline functions B(z), have been constructed for any arbitrary . Finally, it has been
shown that, the following PR relation is satisfied. P ( z ) A( z 1 ) Q ( z )B ( z 1 ) 1 P ( z ) A( z 1 ) Q ( z ) B ( z 1 ) 0
(6)
Fig. 1 shows the complete PR E-spline wavelet system.
s B c, s s (1)
(3)
Quite recently [4], the complete perfect reconstruction E-splines wavelet family, has been constructed for any arbitrary choice of . First, the
2. Mathematical Background
lower ones, i.e.:
. ..
0 m Bm 1 2 m Bm 2 0
E-spline polynomial, i.e. Bm ,L (t ) Bm .
The Exponential m order spline polynomial B m ( t ) , is constructed as m successive convolution of
0
The solution of this system results in exact interpolation for any specified . The L-interpolated E-spline polynomial is defined by inserting (L – 1) equi-spaced points between every two knots of the
Bm (t )
th
...
(3) Fig. 1
Exponential spline PR wavelet family.
1499
Image Super Resolution and Enhancement Using E-spline
3. E-spine Perfect Reconstruction System Since we already proved that P(z)P(z1)R(z) P( z)P(z1)R(z) R(z2. )
P(z1)R(z) z(2m1) P(z) 2m1 P ( z ) z R(z)P(z1) 1 R(z2 ) R(z2 ) P(z)A(z1) Q(z)B(z1) 1
i.e.
Moreover, since R(z) P ( z 1 ) P ( z ) R(z2 ) P ( z ) Q ( z )) B ( z 1 ) R ( z ) P ( z 1 ) R(z2 ) P ( z )) A ( z 1 )
i .e .
P ( z )) A ( z 1 ) Q ( z )) B ( z 1 ) 0
Thus, the filter banks P(z), Q(z), A(z-1)&B(z-1) constitutes a PR system
4. E-Spline Sharpened Interpolators To construct a super resolution image SR, from a given single low resolution image LR [15-18]; one normally estimates the missing high frequency details from the given low resolution image using the information contained in the whole LR image. This is achieved as follows: (1) The LR image, (XL), is decomposed into patches of N N non-overlapping blocks. (2) Each of these blocks scans the whole LR image to get the most M similar candidates to that block. Similarity is measured by those having the least distance in the first 4 moment. (3) Having obtained the M candidates, filter each of these candidates by a “Laplacian” filter to estimate its high
frequency
energy.
Denote
these
filtered
components by Xi, i = 1, ..., m. To further estimate the high frequency content; decompose Xi by 1 level wavelet decomposition. Denote the energy in the detail bank of the decomposed Xi by Ei.
(4) An updated image with improved high frequency content is obtained as Ei Xup = ( )Xi, Es Ei . These Xup images Es
constitute an N N patches of the updated high frequency content of the LR image XLUP. (5) For an arbitrary α of the E-spline function, interpolate Xn XL XLUP to the desired interpolation level. The parameters α’s and μ’s are chosen to sharpen the interpolated image as follows: For an arbitrary α and μ, interpolate Xn to the desired scale to yield an interpolated image Yp. Evaluate its high frequency energy as described in step 4 i.e.: simply by filtering the interpolated image by a “Laplacian” filter. Denote its energy by Ed. Using any unconstrained minimization technique, determine α and μ that minimize the following 2
objective function Y Yp 1 / Eg , where Y is the classical B-spline interpolated image of XL. The second step of sharpening the interpolated image Yp, is achieved by boosting the energy of its finest detail wavelet decomposition as follows: (1) Decompose Yp by 1 level wavelet decomposition. (2) Filter its horizontal and vertical sub-bands by an arbitrary M × M block, and add the filtered output to the diagonal sub-band. (3) Using the updated detail bank, estimate the reconstructed image Yr. Finally, in order to further boost high frequency detail energy and remove any dependence between the initial interpolated B-spline image Y and the optimized E-spline one Yp, these 2 images were processed by a FICA (Fast ICA) algorithm [19]. Table 1 shows the final enhanced PSNR in dB of some
Table 1 B-spl. Y E-spl. Yp Sharp. E-spl Ysh ICA
Cameraman 24.28 24.91 25.17 25.20
Lena 27.22 28.01 28.44 28.33
Mandril 25.06 25.41 25.64 25.60
Boats 25.51 26.00 26.12 26.15
1500
Image Super Resolution and Enhancement Using E-spline
standard Matlab images using regular B-splines, E-splines (first step only without sharpening), E-splines (with sharpening) and E-splines (with sharpening and ICA boosting). As a final example, consider the 1074 Ă— 810 old man image as well as 274 Ă— 184 girl image. Both images were decimated using L = 2, to quarter of its
size. Figs. 2 and 3, compare the performance of the classical B-spline based interpolators with the proposed sharpened super resolution ICA E-spline based interpolator. The resulting PSNR for old man image for both interpolators are 32.67 dB, 33.31 dB, respectively, whereas in the girl image case, the resulting PSNR are 31.4604 dB, 33.94 dB, respectively.
Decimated Image
Cubic Bspline Case Sharpened SR. Espline Case Fig. 2 Left: Decimated old man image using L = 2. Right: The performance of old man image case using the classical B-spline and the proposed sharpened SR E-spline interpolation. The resulting PSNR are 32.67 dB, 33.31 dB, respectively.
Decimated Image
Cubic Bspline Case
Sharpened SR. Espline Case
Fig. 3 Left: Decimated girl image using L = 2. Right: The performance of girl image case using the classical B-spline and the proposed sharpened SR E-spline interpolation. The resulting PSNR are 31.4604 dB, 33.94 dB, respectively.
Image Super Resolution and Enhancement Using E-spline
5. Conclusions As a result of the extra degrees of freedom possessed by E-splines, they enjoy a higher level of energy concentration than their B-spline counterpart. Hence, when its parameters are correctly chosen, they can provide enhanced performance in different image de-noising as well as zooming applications. Simulation results have verified that they improved denoising performance and its ability of constructing interpolators with an even sharper image than those available in the existing methods. This is a crucial problem in SR system design.
References [1] [2]
[3]
[4]
[5]
[6]
[7]
M. Unser, Splines—A perfect fit for signal and image Processing, Signal Processing Magazine 16 (1999) 22-38. M. Unser, A. Aldroubi, M. Eden, B-Spline Signal Processing: Part I—Theory, IEEE Transactions on Signal Processing 41 (1993) 821-833. M. Unser, A. Aldroubi, M. Eden, B-Spline signal processing: Part II—Efficiency design and applications, IEEE Transactions on Signal Processing 41 (1993) 834-848. M.F. Fahmy, G. Fahmy, O.F. Fahmy, B-splines wavelets for signal de-noising and image compression, Journal of Signal, Image and Video Processing 5 (2011) 141-153. M. Unser, T. Blu, Cardinal exponential splines: Part I—Theory and filtering algorithms, IEEE Transactions on Signal Processing 53 (2005) 1425-1436. M. Unser, Cardinal Exponential splines: Part II—Think analog act digital, IEEE Transactions on Signal Processing 53 (2005) 1439-1449. M.F Fahmy, G. Fahmy, Image compression using exponential B-spline functions, in: 29th National Radio Science Conference (NRSC), Cairo, Egypt, Apr. 10-12,
[8]
[9]
[10]
[11] [12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
1501
2012. G.F. Fahmy, M.F. Fahmy, E-splines in image de-noising applications, in: EUROCON, University of Zagreb, Croatia, Jul. 1-4, 2013. G. Bellettini, V. Caselles, M. Novaga, The total variation flow in RN, Journal of Differential Equations 184 (2002) 475-525. C.R. Vogel, M.E. Oman, A Fast, Robust algorithm for total variation based reconstruction of noisy, blurred image, IEEE Trans. Image Processing 7 (1998) 813-824. S. Mallat, A Wavelet Tour of Signal Processing, Academic Press, 2009, pp. 535-610 (Chapter 11). M.F. Fahmy, G.A. Raheem, O.M. Sayed, O.M. Fahmy, A new total variation based image denoising & deblurring technique, in: 2013 EUROCON, Zagreb, Croatia, Jul. 1-4, 2013. S.H. Chan, R. Khoshabeh, K.B. Gibson, P.E.Gill, T.Q. Nguyen, An augmented Lagrangian method for video restoration, in: ICASSP, Prague, Czech, May, 2011. S.H. Chan, R. Khoshabeh, K.B. Gibson, P.E.Gill, T.Q. Nguyen, An augmented Lagrangian method for total variation video restoration, IEEE Trans. on Image Proc. 20 (2011) 3097-3111. H. He, W.C. Siu, Single image super-resolution using gaussian process regression, in: Computer Vision and Pattern Recognition (CVPR), Colorado Springs, USA, Jun. 20-25, 2011. Y. Tang, P. Yan, Y. Yuan, X.L. Li, Single image super-resolution via local learning, International Journal of Machine Learning & Cybernetics 2 (2011) 15-23. W.L. Freeman, T.R. Jones, E.C. Pasztor, Example—based super-resolution, IEEE Computer Graphics and Applications 22 (2002) 56-63. K.I. Kim, Y. Kwon, Example-based learning for single image super resolution and JPEG artifact removal, Tech. report, Max Planck Institute for Biological Cybernetics, Aug. 2008. A. Hypavarinen, J. Karahunen, E. Joha, Independent Component Analysis, John Wiley and Sons Ltd., 2011.
Journal of Communication and Computer 10 (2013) 1502-1506
New Cloud Consolidation Architecture for Electrical Energy Consumption Management Nawfal Madani, Adil Lebbat, Saida Tallal and Hicham Medromi Systems Architecture Team, Department of Computer Information System, ENSEM, Hassan II University, Casablanca 20150, Morocco
Received: October 08, 2013 / Accepted: November 10, 2013 / Published: December 31, 2013. Abstract: Cloud computing is taking more extensive space in the research field. Cloud architectures will need to worry about energy in its various forms, to be profitable, on the one hand, and comply with environmental constraints (energy consumption and CO2 emission) on the other hand. VM (virtual machines) consolidation (VMs in this document), among other techniques, must take into account the consumption of electrical energy, for example, while providing a level of performance that meets the requirements of SLAs (service level agreements). In our work, we focus on an architecture configuration to manage virtual machines in a data center, in order to optimize the consumption of energy, and meet SLAs’s constraints at the same time, by grafting a tracing component of the multiple consolidation plans that leads to an optimal configuration to finally give the order of the migration machinery to a minimum number of servers switched on, knowing that VMs (virtual machines) that coexist in the same server, are at risk of congestion and interference. Key words: Cloud computing, virtualization, power consumption, congestion, VMs consolidation.
1. Introductionď€ The pushing demand on the Internet and its immense use throughout the world have made its use change from a fairly traditional computing to distributed computing, replaced in the trodden by grid computing and the Cloud. Cloud computing has recently emerged as a new paradigm for hosting and providing services through the Internet, allowing access to the practical application of networks, servers, mass storage and specific services with a minimum of effort for both the supplier and the end user. In this model, resources (CPU, storage, RAM...) can be provided for a general use where they can be used or released through the Internet. In a Cloud environment, the traditional role of a provider is now split into two: Architecture provider who manages the Corresponding author: Nawfal Madani, Ph.D. student, research fields: cloud computing, virtualization and security. E-mail: nawfal.madani@gmail.com.
platform and service provider who rent the first’s platforms to serve the end user. In its emergence, the Cloud had a huge impact on the world of information technology industry; in fact, it provided some interesting features for IT business in the world, as we will be able to demonstrate: (1) No investment fund: Cloud computing is based on a pay-per-use pricing model, in which a service provider does not need to invest in a platform to start generating income, it can simply rent the resources, it needs and pay for the use of these resources. (2) Risk and maintenance costs reduction: By outsourcing its infrastructure, a service provider moves its risk (hardware problems) to infrastructure supplier, who should have the skills and expertise necessary to manage such risks; hence, the service provider will save maintenance and training costs. (3) Highly scalable platforms: Platform providers can consolidate large numbers of resources in data centers, making them easily accessible. From this
New Cloud Consolidation Architecture for Electrical Energy Consumption Management
moment, the service provider can expand its service to a larger scale, with the aim to solve the problems of sudden load increase (flash crowd effect). (4) Easy Access: The services available through the Cloud are generally based on the Web. Therefore, they are available for any device with an Internet connection (computers, mobile phones...). (5) Lower operating costs: The Cloud resources can be rapidly allocated or released on demand. Consequently, a service provider does not need to be supplied according to the peak load. This generates savings since these resources can be released to lower operational costs at the time of falling demand. However, although the Cloud concept shows significant opportunities for the IT industry, it denotes also important issues that must be carefully identified and studied to be resolved. This paper will be organized as follows: Section 2 presents a background of the consolidation of VMs and works related to optimal management of energy, in Section 3 we propose a new architecture for the VMs consolidation and conclude by outlining future prospects in final section.
2. Energy Management Problem in the Cloud Although the Cloud concept is widely adopted by IT industry, research in this area is still in its early stages. Several challenges have not yet been fully addressed, while new are emerging from applications industry. In the next section, we summarize some research issues relating to energy management in the world of Cloud computing. Improving energy efficiency is still a major problem for the Cloud. Powering and cooling machines cost is around 53% of the operational budget for data centers [1]. Therefore, infrastructure providers are under enormous pressure to reduce energy consumption. Since their aim is not only to reduce the cost of energy in data centers [2], but also to obey government regulations and the new concept
1503
of green computing [3]. Server consolidation would be an optimal approach for the maximum use of resources with a minimum consumption of energy in a Cloud environment. Live virtual machines migration is often used to consolidate VMs running on multiple low accessed servers to a single server, in order to put other servers in power saving status. It is also noted that the dependencies between VMs in terms of intercommunication is still a challenge to overcome [4]. However, consolidation is not expected to have an impact on applications performances. Many literature works have tried to capture and exploit the characteristics of shared running VMs pages in order to collocate similar VMs to reduce the total memory footprint [5]. We should also remember that the fact of consolidating to the maximum server resources, shared between VMs, i.e. bandwidth, cache and disk input-output can generate a congestion phenomenon when one of VMs change its footprint [6]. This leads to think about this problem of congestion by observing VMs footprint fluctuations and using this information to efficiently consolidate servers. VMs consolidation raises many problems related to optimized resources management in a way to avoid severe performance degradation due to the coexistence of VMs. Several studies that have addressed these optimization problems, focused, on the one hand, on physical constraints covering the consumption of CPU and shared memory on the local node, and providing load management and number of switched on servers algorithms [7], in addition to the deployment of this algorithms in a real Cloud environment, on the other hand [8].
3. Vms Consolidation in an Infrastructure Virtualization technologies and network architectures provide the basis for the Cloud management solutions. Several contemporary products offer dynamic access to the resources of data
1504
New Cloud Consolidation Architecture for Electrical Energy Consumption Management
centers. Some open-source solutions exist in the market; however, they have a gap in dynamic services delivery mechanisms and architecture management evolution. Management platform must indeed provide performance indicators which provide information on the costs and the VMs network, as well as energy consumption. It must also be equipped with a decision mechanism for optimal VMs, a decision that turns out to be complex in terms of computing speed as it implements multi-criteria searches problems, alongside with space optimizing, servers computing speed limits and SLAs constraints. To cope with these constraints, a Cloud platform generally follows a model based on four management stages (0). Step 1: Collect monitoring data; Step2: Exploit these data to calculate a better VMs placement model; Step3: Draw a plan to migrate virtual machines; Step 4: Apply the proposed moves. The platform model we proposing is composed of a controlling monitor, a calculation component, a better plan tracing component and a migration mechanism that communicate in a pipeline process, where one part’s output information is the next’s input. Calculation component mainly contains three sub-modules that command separately: server resources, network traffics and electricity (0). At the end of the best plan tracing component process, the placement mechanism is to implement these recommendations.
3.2 Server RAM and CPU VMs that coexist in the same server are sharing RAM and CPU. SLAs come into play to prevent a resources supply, which is not allowed. The main goal of the consolidation is the ability to run as many virtual machines as possible onto a single physical server, avoiding the unavailability of resources. In an operating system dedicated to a general use, when two tasks start to run, they come into competition for the underlying jobs on computational resources. The part that is assigned to each task can be divided according to their needs and time requirements. 3.3 Power Consumption Power consumption is very important for two main
3.1 Data Center Network
reasons. On the one hand is the point of environmental,
It is very important to understand that bandwidth
Fig. 1
and latency are the two most important attributes in our context of energy management, while these two are different in terms of importance, depending on the type of the provided service. Service MapReduce [9] describes, for example, the bandwidth to be more important since it reduces the time of data exchange, interactive services, in the meanwhile, prefer a minimal latency in order to avoid edge effects during interactive use [10]. In fact, if we are in a scenario of a given data center links congestion, where the center does not have a separate network dedicated to VMs migration, the introduction of a live migration process will probably lead to a general
on the other hand is economic. To implement this
Controlling monitor (performance indicators)
Computation component (CPU, RAM, Power)
Migration mechanism
Best plan tracing component
Conception of the proposed platform.
New Cloud Consolidation Architecture for Electrical Energy Consumption Management
Fig. 2
1505
The platform architecture.
management module, the Cloud platform must have access to load and power consumption information to achieve an effective energy management policy. In order to achieve an optimal energy control, some techniques inform on servers electrical consumption thresholds, based on the computation load of each. Consumption rate changes depending on computing load, which sometime makes placement algorithms impossible to implement due to lack of resources. It should be noted that VMs introduce a notion of load variation depending on the type of service provided. In peak hours, the load increases, and prohibits consolidation, while during the night hours, load on servers decreases, allowing an efficient consolidation of underused servers. Dynamic VMs allocation explains the importance of live migration to increase the energy management in a Cloud environment.
applications. These applications running on VMs are subject to several factors, namely the coexistence of VMs and virtualization. In order to optimize energy management, we have proposed a shared management architecture based on a modular open-source solution. In this model, our proposed architecture is feasible through the modules of the OpenStack solution, which will be our future work, but we must, however, be vigilant to avoid performance degradation. At this stage of research, we are interested in measures of the effects of consolidation on performance in a context governed by SLAs, in order to optimize the choice of placement and VMs migration algorithms, while avoiding side effects due to changes in the footprint of VMs.
References [1]
4. Conclusions In conclusion, the emergence of shared infrastructure such as the Cloud has imposed significant challenges for both providers and
[2]
J. Hamilton, Cooperative expendable micro-slice servers (CEMS): Low cost, low power ervers for Internet-scale services, in: 4th Biennial Conference on Innovative Data Systems Research (CIDR), Asilomar, California, January 4-7, 2009. Gartner Estimates ICT Industry Accounts for 2 Percent of Global CO2 Emissions [Online], April 26, 2007, http://www.gartner.com/it/page.jsp?id=503867, 2011.
1506 [3] [4]
[5]
[6]
New Cloud Consolidation Architecture for Electrical Energy Consumption Management S. Murugesan, Harnessing green IT: Principles and practices, IEEE IT Professional 10 (2008) 24-33. X. Meng, Improving the scalability of data center networks with traffic-aware virtual machine placement, in: 2010 Proceedings IEEE INFOCOM, San Diego, CA, March 14-19, 2010. T. Wood, G. Tarasuk-Levin, P. Shenoy, P. Desnoyers, E. Cecchet, M.D. Corner, Memory buddies: Exploiting page sharing for smart colocation in virtualized data centers, in: Proc. of the ACM SIGPLAN/SIGOPS International Conference on Virtual Execution Environments, Washington, DC, March 11-13, 2009. P. Padala, K.Y. Hou, Automated control of multiple virtualized resources, in: Proceedings of the 4th ACM European Conference on Computer systems, New York, NY, 2009, pp. 13-26.
[7]
S. Lee, R. Panigrahy, V. Prabhakaran, V. Ramasubramanian, K. Talwar, L. Uyeda, U. Wieder, Validating Heuristics for Virtual Machines Consolidation, Technical Report, 2011. [8] C. Isci, J.E. Hanson, I. Whalley, M. Steinder, J.O. Kephart, Runtime demand estimation for effective dynamic resource management, in: Proc. of the IEEE Network Operations and Management Symposium, NOMS, Osaka, Japan, 2010. [9] J. Dean, S. Ghemawat, MapReduce: Simplified data processing on large clusters, ACM Communications 51 (2008) 107-113. [10] X. Meng, V. Pappas, L. Zhang, Improving the scalability of data center networks with traffic-aware virtual machine placement, in: Proc. of the 29th Conference on Information Communications (INFOCOM’10), San Diego, USA, 2010.
Journal of Communication and Computer 10 (2013) 1507-1521
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real Indoor and Outdoor Location in WSN Mauro Rodrigo Larrat Frota e Silva, Leomário Silva Machado and Dionne Cavalcante Monteiro Exact and Natural Sciences Institute, Faculty of Computer Science, Federal University of Para, Belem-PA 66075-110, Brazil Received: October 23, 2013 / Accepted: November 20, 2013 / Published: December 31, 2013. Abstract: WSN (wireless sensors networks) is a promising research area which has been in center of attention in many applications on telecommunications. Despite of many existing applications, a bunch of algorithms have been created or applied to solve different issues surrounding WSN. An attractive subtopic on research area is Localization Algorithms. Due to the countless applications, it is almost impossible to list all the algorithms applicable to solve sensor’s location problem in WSN over distinguished parameters associated to diverse environments. This paper evaluates a geometrical algorithm, an instance based algorithm and a function approximator algorithm, having the RSSI (received signal strength indicator) as metric to estimate planar coordinates in an indoor and outdoor environment using a WSN based on IRIS mote. The analysis of the WSN is constructed over statistical data obtained from empirical experiments and the observed characteristics of the algorithms. We also estimate the performance for different parameters configurations applied to the algorithms for both indoor and outdoor environment. Also, some comments about the tradeoff between the accuracy and the performance of the algorithms are made objecting to. Also, some objections about the tradeoff between the accuracy and the performance of the algorithms are made when relevant. Key words: Wireless sensors network, RSSI, artificial neural network, Location Algorithms, lateration, KNN.
1. Introduction WSN (wireless sensor network) is a type of ad hoc network formed by small low-cost devices (or wireless sensors) capable of interacting together and perform data acquisition, preprocessing and data transmission from the environment to which they are inserted. Such devices have the restriction on energy efficiency, which consequently affects their processing speed, the amount of storage, the limit of the signal transmission distance and complexity of the protocols implemented. These limitations are suppressed by applying sophisticated algorithms techniques and the use of devices that include the latest hardware technologies, Corresponding author: Mauro (aka Mauro) Rodrigo Larrat Frota e Silva, M.Sc., lecturer, research fields: wireless sensors networks and intelligent systems. E-mail: maurolarrat@gmail.com.
seeking the balance between performance and consumption of computational resources. The joint use of a large number of these devices in a network provide many applications in areas such as industrial, agricultural, military, medical and scientific. This is mainly due to the characteristics of scalability and fault tolerance of such networks. Localization is one of the fundamental challenges in WSN owing to randomness inherent to device’s deployment and the need for efficient data routing. Localization is necessary to one device knows the position of another device on the network. Through a good location system, these devices can save energy consumed during transmission and reception of data, thereby increasing the lifetime of the network [1]. Regard the use of GPS for localization, unless its god accuracy, there are some network and factory
1508 Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real Indoor and Outdoor Location in WSN
constraints which can be summarized as follows: Obstacles should not be in the line of sight of the signal; The power consumption is higher than that required for wireless sensors; The deployment costs of a GPS network is much higher than the cost of deploying using wireless sensors, even if it has a few more devices in the second than in the first; The size of the GPS is bigger than the size of the most usable wireless sensors commonly used in applications. There are two main localization approaches in WSN: range-free and rage-based [2], where each one can be implemented through many kind of methods [3-7]. For low cost implementation and avoidance of additional hardware, we chose a range-based method using the RSSI (received signal strength indicator) that measure the signal intensity from RF (radio frequency) between devices on the network [8]. The RSSI is a metric that are implemented in the 802.11 radio family which need at least three devices with known positions (beacon devices) to estimate the location of any other devices (sensor nodes) in the network range. The formula that can describe RSSI is shown in Eqs. (1) and (2): – 10
log 10
(1) (2)
The values of A are calculated considering the RSSI in one meter of distance. We calculated A = 38 in both environments. The n (path loss) is the attenuation introduced by the propagation environment of a signal that can be calculated through a rearrangement in Eq. (1). In this paper, we consider a non-obstructed area in the experiment, so we can use the simple RSSI model as shown in Eq. (1), which does not consider the signal fluctuations due to obstacles or the orientation of devices on the network during the communication [9]. Having the RSSI data stored, we apply those data to three kinds of algorithms to estimate the location of any
sensor node inside the area of the experiment. We calculate the matrix location of each sensor node using Lateration [10], estimate the coordinates of sensor nodes as weights in the arithmetic RSSI average of its KNN (K-Nearest Neighbors) [11] and we apply the RSSI as inputs of an ANN (artificial neural network). This paper presents an experimental comparative between these three algorithms in an indoor and non-obstructed environment. Thereafter, we make a second experiment in an augmented and non-obstructed outdoor area to analyze the effects of scalability and changes in the environment over the algorithms. Section 2 shows some recent related works in the same area of research. Section 3 shows the methodology of our experiments. Section 4 shows our tests and Section 5 shows ours achieved results. We finalize in Section 6 with some considerations about this work.
2. Related Work This session introduces some researches about performance evaluation in WSN, highlighting comparisons between localization techniques in WSNs. Ref. [12] performs a comparative study among the techniques Lateration, Min-Max, NN (nearest neighbor) and KNN in an experimental indoor environment 3 m × 3 m using XBEE modules. The article shows Lateration as the technique with minor error, and then KNN technique. Ref. [12] shows a study that compares the Location Algorithms NLR (non-linear regression), INLR (iterative non-linear regression), LS (least squares), RANSAC (random sample consensus) and ToM (trilaterate on minima). Data were collected in real environments in an area of 550 m2. The radios use UWB (IEEE 802.15.4a) in their PHY layer for transmitting and receive signals. The INLR technique gets smaller estimation errors in location. In Ref. [13], a comparative work in real environment is performed among several ANN based on methods
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real 1509 Indoor and Outdoor Location in WSN
such as RBF (radial basis function), MLP (multi-layer perceptron), RNN (recurrent neural networks), PV (position-velocity), PVA (position-velocity-acceleration) and RRBF (reduced radial basis function). These methods evaluate the location errors in centimeters using these techniques in a 3 m × 3 m indoor environment. However, they do not compare the performance between some of the ANN method with the common location methods applied in real environment. In Ref. [14], they perform, in a real environment, a comparison between a MLP neural network model and two Kalman Filter models, namely PV and PVA techniques. The environment measure 3 m × 3 m, marked in grid spacing of 0.30 m. Four beacon nodes are positioned on the four vertices of the square. The mobile nodes are placed on each intersection of the grid to collect the data. The experimental results indicate that among the three Localization Algorithms, the MLP neural network has the best performance in terms of accuracy beyond memory usage, and computation complexity. However, there is a potential retraining or redesign cost associated with the use of neural networks which is not associated with the Kalman Filter. This work also compares neural networks only, not performing tests with mathematical techniques. Ref. [15] is performed a simulation in a environment measuring 200 m × 200 m, where they implement an ANN with Levenberg-Marquardt Algorithm using an error minimization function, the simulation shows that the location accuracy can be better with the increase of grid sensors density and APs. The work does not present the benchmark among the techniques analyzed, disadvantaging a preview comparative with better clarity. In Ref. [16], Lateration and Min-Max algorithms are simulated together with the sum of the distances, DV-Hop and Euclidean algorithms. The simulation performed describes an ideal environment without discoursing about the effect of path loss or variable
depending on the Lateration. In Ref. [17], a new ANN method known as LNNE (localization neural network ensembles) is compared with DV-Hop and a LSNN (localization signal neural network). This work employs a simple scenario controlled by simulation. Ref. [18] compares three methods of localization namely OML (optimal multilateration), SBT (sub-optimal blind trilateration) and GDOP (geometric dilution of precision). They provide the benchmark in terms of achievable accuracy. However, they only compare the performance among Lateration variants, not comparing the techniques of other nature. This paper proposes to conduct a performance evaluation in a real indoor and outdoor environment to compare two mathematical techniques (Lateration and KNN) with an intelligent technique (ANN) as a function of the error of the estimated accuracy for each technique.
3. Methodology This section presents the methodology of the data acquisition and its characteristics as well as the environment used during the experiments. Then, we describe the parameters of algorithms used in this evaluation. 3.1 Characteristics of Indoor and Environment and the Data Acquisition
Outdoor
The test in an indoor environment was conducted in a non-obstructed area measuring 12 m × 12 m as seen in Fig. 1. The beacon nodes were positioned in the corners of the scenario. The beacon nodes collected 50 samples (RSSI) from each of the 82 points representing the sensors node in the internal area to acquire the data used in the algorithms, totalizing in 4.510 samples. In outdoor environment, the test was performed in a non-obstructed opened field measuring 20 m × 22 m as seen in Fig. 2. Here, the beacon nodes collected 1,000 samples from each of the 21 points inside the area, totalizing in 21.000 samples.
1510 Perform mance Evaluation of Lateration, KNN and a Artificial Neural Networks Techniq ques Applied to Real ndoor and Ou utdoor Location in WSN In
use as inputs to the algorithm ms. Two outp puts representt the coordinates which the seensor nodes were w locatedd during the data d acquissition. Thee statisticall chaaracteristics of o the samplles which were w acquiredd from m the beaconn nodes in the indoor env vironment aree shown in the Tabble 1. Those characteristiccs referred too the outdoor enviironment are shown in Tab ble 2. The T statisticaal characterisstic of thesee coordinatess colllected in the indoor i and in the outdoor environments e s are shown (norm malized) in Taables 3 and 4 respectively.. Fig. 1 Poin nts in which the RSSI signals are colleected through the sensor nodes in n an indoor areea.
Tab ble 1 Sampless characteristiccs in indoor environment. Staatistics Min n. value 1st. quartile Meedian Meean 3rd d. quartile Maax. value
Beeacon 1 0.000 2.000 3.000 2.7743 4.000 7.000
Beacoon 2 0.00 3.00 4.00 4.1999 6.00 9.00
Beacon 3 0.00 2.00 4.00 3.802 5.00 8.00
Beacon 4 0.00 2.00 3.00 3.221 4.00 8.00
Tab ble 2 Sampless characteristiccs in outdoor environment.
Fig. 2 Poin nts in which the RSSI signals are colleected through the sensor nodes in n an outdoor arrea.
We chosee 75% of sam mples for trainning and 25% % for testing the algorithms a in each case. These T percentaages on samples for f training and a testing weere chosen frreely as a rule of o thumb. The T samples were randoomly shuffled beffore this division, to ensuure that the ANN A Algorithm did d not preseent trends in estimation. This T shuffling does not affect the t results in the KNN as well as in the Latteration. The tempperature oscilllated 4 째C inn the indoor and outdoor envvironment foor the time we w executed the experiment. The humiddity oscillateed 5% in both b environmentts. The timee interval we w executed the experimentss was 6 hours approximateely. We mappped each 4-tuuple of samplles taken at same s time from thhe four beacoon nodes as four attributees to
Staatistics Min n. value 1st. quartile Meedian Meean 3rd d. quartile Maax. value
Beeacon 1 0.000 3.000 11.00 11.36 166.00 277.00
Beacoon 2 0.00 6.00 11.000 11.199 15.000 25.000
Beacon 3 0.00 4.00 10.00 10.87 17.00 28.00
Beacon 4 0.00 8.00 12.00 12.08 18.00 24.00
Tab ble 3 Characcteristics of coordinates calculated from m sam mples in an indooor environmeent. Staatistics Min n. value 1st. quartile Meedian Meean 3rd d. quartile Maax. value
X 0.00 0.50 0.6667 0.6235 0.8333 1.00
Y 0.00 0.25 0.50 0.5179 9 0.8333 3 1.00
ble 4 Characcteristics of coordinates calculated from m Tab sam mples in an outd door environm ment. Staatistics Min n. value 1st. quartile Meedian Meean 3rd d. quartile Maax. value
X 0.00 0.25 0.50 0.50 0.75 1.00
Y 0.00 0.2292 2 0.4583 3 0.4583 3 0.6875 5 0.9167 7
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real 1511 Indoor and Outdoor Location in WSN
The statistical information shown in Tables 1-4 were generated using RStudio Software [19]. These information were generated using the dataset from each environment, composed of samples from each beacon node in the corner (red points in Figs. 1 and 2) and the referred x and y coordinates of the sensor devices (blue points in Figs. 1 and 2). With these statistical information, one can see that there is no outliers and extremes values in all the samples, so the KNN and Lateration will not be impaired by the ability of the ANN in dealing with such unwanted values. 3.2 Hardware, Middleware and Software The sensor used is the IRIS motes of MEMSIC, which implements a 2.4 to 2.48 GHz IEEE 802.15.4 radio standard with a maximum rate of 250 Kbps. It can transmit data at distances up to 500 m in line of sight. The IRIS Hardware is based on XM2110CB radio and ATmega1281 CPU, both from ATMEL. To support the WSN node gateway, it used a MIB520, which consists of a USB adapter that enables two serial communications, one for control and one for data. The IRIS runs the TinyOS operating system with open source BSD license [20]. A java based application extracts the data in real time and records it in a file with the following fields: actual localization, node identifier and the RSSI value obtained to target node. 3.3 Lateration Algorithm Lateration in an algorithm that calculates the location of one point in a space based on the known position of other points. In this paper, four points are considered for localization. Based on Eqs. (1) and (2), for each message received by an i-th beacon node, the Euclidean distance is determined by Eq. (3). (3) where, x and y are coordinate values related to the sensor node, and are coordinate values related to the i-th beacon node, is the distance between the sensor node and the i-th beacon node. For each k visible anchor nodes, there will be k
distances and k equations as a function of x and y coordinates. They can be summarized in matrix form as shown in Eq. (4). B is the estimated location vector from coordinates x and y. (4) where, (5) (6) and, 2 x 2 x
2 y
x x
2 y
y y
(7)
3.4 K-Nearest Neighbor Algorithm KNN is one of the most popular instance based methods of classification due to its simplicity and reasonably accuracy. It requires assembling a specific model and has shown good performance for the classification of various types of data [21]. For sensor localization with the KNN, the RSSI taken from the beacon messages were assumed to be the instances for classification. For implementation, the KNN function from caret library, which can be downloaded in the RStudio Software. For each of the environments, we train the KNN Algorithm using ten folds in a cross-validation method over all the instances collected in each experiment. We repeat the cross-validation method five times and the KNN Algorithm minimizes the cost of a RMSE (root mean squared error) function. Also, we run the KNN Algorithm for k = 1 to k = 10, where k is the number of neighbors, and plot show the results as RMSE. Figs. 3 and 4 show the RMSE for x and y coordinates respectively for the values of k, related to the indoor environment. Figs. 5 and 6 show the RMSE for x and y coordinates respectively for the values of k, related to the outdoor environment.
1512 Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real Indoor and Outdoor Location in WSN
Fig. 3 RMSE for x coordinate showed for all k. Related to indoor environment.
Fig. 4 RMSE for y coordinate showed for all k. Related to indoor environment.
Fig. 5 RMSE for x coordinate showed for all k. Related to outdoor environment.
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real 1513 Indoor and Outdoor Location in WSN
Fig. 6 RMSE for y coordinate shown for all k. Related to outdoor environment.
Despite of the lines tendency in Figs. 3 and 4, the x coordinate proved to be better classified than y coordinate for all values of k. This behavior reflects the statistics of the data shown in Tables 3 and 4 previously and it occurs in all ensuing algorithms. For further algorithm comparison, we chose k = 4 for both indoor and outdoor environments. 3.5 Artificial Neural Network ANN are good function approximators as well as good classifiers that can be thought as a set of interconnected mathematical models based on the practical characteristics of mammalian neurons. Here, the RSSI are used as inputs to train and validate the network. The x and y coordinates are used as network outputs to estimate. Multilayer Feedfoward Perceptron Architecture, trained by the Backpropagation Algorithm, was our choice for ANN implementation. The application was developed using the MLP function from RSNNS library in the RStudio to run the network. The network has two hidden layers. The ANN uses ten folds in a cross-validation method repeated five times, as we did in the KNN experiment, so the RMSE is taken as cost function too. The weights decay is set to 0, 0.001 and 0.1 for each configuration. The weights decayment is necessary due to the unnecessary increase of the weights during the tanning update, so it
is a factor of the last weight that is subtracted from the new weight value during the Backpropagation operation. The number of neurons in each hidden layer was from 1 to 6 neurons. Figs. 7 and 8 show the RMSE for x and y coordinates respectively for all ANN parameters, related to the indoor environment. Figs. 9 and 10 show the RMSE for and coordinates respectively for all ANN parameters, related to the outdoor environment. We can see in the figures above that the final values used with the models for the indoor and the outdoor were two hidden layers with 6 hidden units and decay equal to 0.1.
4. Results The absolute values of positioning errors in the coordinates X (Xerror), Y (Yerror) and the Le (Location error) are used as a metric for comparing the techniques. The position error is the distance between the actual point and the estimated value as shown in Eqs. (8)-(10). | | (8) | | (9) (10) To compare the results between Lateration, KNN and ANN techniques, we use the Eqs (11) and (12) to
1514 Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real Indoor and Outdoor Location in WSN
Fig. 7 RMSE for x coordinate showed for all parameters set in ANN. Related to indoor environment.
Fig. 8 RMSE for y coordinate showed for all parameters set in ANN. Related to indoor environment.
Fig. 9 RMSE for x coordinate showed for all parameters set in ANN. Related to outdoor environment.
Perform mance Evaluation of Lateration, KNN and a Artificial Neural Networks Techniq ques Applied to Real 15155 ndoor and Ou utdoor Location in WSN In
Fig. 10 RMS SE for y coordiinate showed for f all parametters set in ANN N. Related to outdoor o environ nment.
generate the graphs of thee benchmarkss.
4.1 Benchmark in i the Indoor Environmentt
100
(11)
100
(12)
where, : Benchmaark of ANN versus v Laterattion; : Benchm mark of ANN versus v KNN;; : Location error e obtaineed from ANN N in meters; : Location errror obtainedd from Lateraation in meters; : Location error e obtaineed from KNN N in meters.
In n Fig. 11, wee can observee the compariison betweenn AN NN and Laterration, in peercentage, wh here positivee valu ues indicate that ANN has better results thann Lateration. Sim milarly, Fig. 12 shows a comparisonn betw ween ANN and a KNN. T To illustrate these t graphs,, only y 50 points were selecteed from the total t used too estiimate the locaation. The T ANN achhieved better rresults in the location withh aveerage 232.43% % better thann Lateration and a 470.45% % bettter than KNN N. Table 5 shows the percentage p off casees in which eaach techniquee has the best performancee for the aforemenntioned metriccs.
Fig. 11 Perccentage comparrison between the ANN and Lateration in an a indoor Envvironment.
1516 Perform mance Evaluation of Lateration, KNN and a Artificial Neural Networks Techniq ques Applied to Real ndoor and Ou utdoor Location in WSN In
door Environm ment. Fig. 12 Perccentage comparrison between the ANN and KNN in an ind
The data in the Table 5 show thhat ANN obttains 76.0784% and a 79.6078% % of the best samples in the x and y coorddinates, resppectively, whhen comparedd to Lateration. In I the compaarison of KNN N and ANN,, the Table 5 shhows that 85.8824% annd 88.6275% % of samples havve better resuults for x annd y coordinaates, respectively. Therefore, the t ANN shoows better ressults when compaared with Latteration and KNN. K When wee consider thhe distance of o the estim mated location andd real locationn, the ANN has h better ressults than Laterattion and KNN N, i.e. the AN NN estimatess the location neaar to the real location in 944.902% compared to Laterationn and 98, 03992% compared to KNN. Table 6 shhows the Avgg (average) annd Var (variannce) of the positiioning error of each technnique. The mean m and variancee of the positioning errorr in the ANN N are lower than thhose calculatted for Lateraation and KNN N. In Tabless 7-9, samplees are organiized by zonees of error. For errrors which are a smaller than t 1 meter,, the ANN shows the best results r with large advanntage because 65.49% of the estimated poositions for thhe x coordinate are a within thhe lower zoone of error, i.e. Xerror < 1 m. m This same behavior is observed o for the t y coordinate, where w 71.76% % of the sampples has Yerrror < 1 m. The Loocation error between the actual point and estimated point p was 488.62% to thhe samples with w location erroor < 1 m. Fig. 13 and a 14 show estimates of some sampples. Each real pooint is illustraated by a squuare connecteed to
t estimatees by dashed line. its three The T Fig. 13 shhows samplees located in the t peripheryy of the t scenario. With this Fiigure, we can n observe thee beh havior of thesee techniques near the bord ders. The Fig.. 14 shows the internal points situated in th he innermostt regiion of the expperiment. 4.2 Benchmark in i the Outdooor Environmeent n Fig. 15, wee can observee the compariison betweenn In AN NN and Laterration. The ppositive valuees show thatt AN NN can morre accuratelyy estimate the t location. Sim milarly, Fig. 16 1 shows a coomparison beetween ANN N and d KNN. To illlustrate thesee graphs, onlly 100 pointss werre selected from f the tottal used to estimate thee locaation. The T ANN achieved betterr results than n Lateration.. In comparison c w KNN, thhe ANN was less accuratee with for samples thatt occurred booth in the testt and trainingg setss. For samplees that not occcur in both seets, the ANN N has so far more accuracy thann KNN. So KNN K shows a trad deoff betweeen memory and accuracy y. Table 100 shows the percenntage of cases in which eaach techniquee has the best performance p for the afo orementionedd mettrics. The T data in thhe Table 10 sshow that AN NN estimatess the location bettter than Laterration does, but b not betterr than n KNN, for all a samples to both coordin nates. We cann also o observe the t location error, considering thee disttance of the estimated e locaation and reall location.
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real 1517 Indoor and Outdoor Location in WSN Table 5 Best performance: ANNxLAT and ANNxKNN. Comparison ANN × LAT ANN × KNN Metric ANN (%) LAT (%) ANN (%) KNN (%) x 76, 0784 23, 921 85, 8824 14, 1176 y 79, 607 20, 392 88, 627 11, 372 Le 94, 902 5, 098 98, 039 1, 9608 Table 6 Mean and variance of the position error in meters for each technique.
Xerror Yerror error distance
LAT Avg Var 1, 76 1, 27 2, 02 1, 34 2, 94 1, 18
KNN Avg Var 2, 68 3, 55 3, 77 4, 68 5, 01 4, 49
ANN Avg Var 0, 80 0, 51 0, 88 0, 88 1, 30 1, 12
Table 7 Quantitative analysis of samples every meter error in x coordinate. Xerror (m) [0, 1) [1, 2) [2, 3) [3, 4) [4, 5) [5, +∞)
LAT (%) 32, 54 26, 66 25, 88 13, 33 1, 56 0
KNN (%) 16, 47 23, 92 16, 07 21, 56 8, 62 13, 33
ANN (%) 65, 49 28, 23 5, 49 0, 78 0 0
Table 8 Quantitative analysis of samples every meter error in y coordinate. Yerror (m) [0, 1) [1, 2) [2, 3) [3, 4) [4, 5) [5, +∞)
LAT (%) 24, 70 22, 35 30, 19 20, 78 1, 96 0
KNN (%) 11, 37 10, 98 12, 94 17, 25 15, 68 31, 76
ANN (%) 71, 76 16, 07 7, 84 2, 35 1, 96 0
Table 9 Quantitative analysis of samples every 1 meter in distance error. Le [0, 1) [1, 2) [2, 3) [3, 4) [4, 5) [5, +∞)
LAT (%) 5, 09 19, 21 22, 74 35, 29 17, 64 0
KNN (%) 1, 96 6, 27 8, 23 14, 50 21, 17 47, 84
ANN (%) 48, 62 29,80 15,29 3, 52 2, 74 0
Table 11 shows the Avg and Var of the positioning error of each technique. In Tables 12 and 13, samples are organized by zones of error. In Figs 17 and 18, we can observe the estimation of these techniques near the borders and near to the center, respectively, related to outdoor environment.
5. Conclusions For the indoor experiment, the use of a neural network model shows satisfactory results in the estimation of a node within the limits of the beacon nodes. For the outdoor environment, KNN shows a better result if the coordinates are stored in memory, however, the ANN can improve the estimation for both known and unknown location. In indoor, the ANN method had lower mean and error variance when compared to the techniques of Lateration and KNN, therefore, having increased stability. However, the
Fig. 13 Real and estimated positions by technique in indoor environment for peripheral points.
1518 Perform mance Evaluation of Lateration, KNN and a Artificial Neural Networks Techniq ques Applied to Real ndoor and Ou utdoor Location in WSN In
Fig. 14 Reall and estimated d positions by technique t in in ndoor environm ment for intern nal points.
Fig. 15 Perccentage comparrison between the ANN and Lateration in an a outdoor envvironment.
Fig. 16 Perccentage comparrison between the ANN and KNN in an outtdoor environm ment.
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real 1519 Indoor and Outdoor Location in WSN Table 10 KNN.
Best performance: ANN × LAT and ANN ×
Comparison metric x y
ANN × LAT ANN (%) LAT (%) 93, 45 6, 34 96, 03 3, 76
ANN × KNN ANN (%) KNN (%) 7, 93 92, 06 49, 80 50, 20
Table 11 Mean and variance of the position error in meters for each technique.
Xerror Yerror
LAT Avg Var 0, 24 0, 00 0, 56 0, 02
KNN Avg Var 0, 03 0, 01 0, 45 0, 00
ANN Avg Var 0, 03 0, 01 0, 44 0, 01
Table 12 Quantitative analysis of samples every meter error in x coordinate. Xerror (m) [0, 1) [1, 2) [2, 3) [3, 4) [4, 5) [5, +∞)
LAT (%) 0 0 0, 40 4, 40 63, 80 30, 40
KNN (%) 93, 90 0, 40 0 0 0, 20 5, 48
ANN (%) 92,81 0 0 0 1, 59 5, 58
Table 13 Quantitative analysis of samples every meter error in y coordinate. Yerror (m) [0, 1) [1, 2) [2, 3) [3, 4) [4, 5) [5, +∞)
LAT (%) 0, 59 0, 19 0, 59 0, 19 0, 19 98, 21
KNN (%) 91, 81 0 0 0 0 8, 18
ANN (%) 57, 79 0 1, 59 0 0, 39 40, 20
number of inputs is fixed and directly related to the number of beacon nodes, resulting in the need to retrain for different amounts of network beacon nodes. The Lateration is a technique for easy adjustment depending on the number of beacon nodes maintaining stability similar to ANN. This technique requires high fidelity n of the environment and the value of A. Otherwise, the error can be severely attenuated and may exceed the expected limits of maximum and minimum values within the intersection bounded by beacon nodes. In this experiment the technique KNN was performed by weighted sum of each coordinate by dividing the sum of their weights. In KNN, estimative tend to be found within the boundaries of the scenario, due to the weighted sum of RSSIs from their neighbors. In Ref. [12], the performance evaluation indicates the Lateration technique has better precision followed by KNN. In this study, both techniques were compared to ANN for location. We observed that the ANN had, in general, a lower error and better precision in locating the target node in all points, as analysis of the mean and variance of the error for both x and y coordinates and location error, causing good reliability in the use of ANN compared to classical techniques for locating wireless sensor networks.
Fig. 17 Real and estimated positions by technique in outdoor environment for peripheral points.
1520 Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real Indoor and Outdoor Location in WSN
Fig. 18 Real and estimated positions by technique in outdoor environment for internal points.
As future work, we intend to perform the target tracking in real time by applying this current artificial neural network to perform this task; test for different amounts of beacons nodes and test this techniques with others hardware.
[7]
[8]
References [1]
[2]
[3]
[4]
[5]
[6]
A. Pal, Localization algorithms in wireless sensor networks: Current approaches and future challenges, Network Protocols and Algorithms 2 (2010) 45-73. Q. Xiao, Range-free and range-based localization of wireless sensor networks, Ph.D. Thesis, The Hong Kong Polytechnic University, 2011. R. Priwgharm, P. Chemtanomwong, A comparative study on indoor localization based on RSSI measurement in wireless sensor network, in: 2011 Eighth International Joint Conference on Computer Science and Software Engineering (JCSSE), Nakhon Pathom, May 11-13, 2011. S. Rajaee, S. AlModarresi, M. Sadeghi, M. Aghabozorgi, Energy efficient localization in wireless ad-hoc sensor networks using probabilistic neural network and Independent Component Analysis, in: International Symposium on Telecommunications, Tehran, 2008. Y. Jieyang, Z. Liang, Enhanced location algorithm with received-signal-strength using fading Kalman filter in wireless sensor networks, in: 2011 International Conference on Computational Problem-Solving (ICCP), Chengdu, China, Oct. 21-23, 2011. H.C. Hsu, J.H. Wen, C.H. Cheng, Y.S. Chen, T.Y. Wu, TOA Estimation with DLC Receivers for IEEE 802.15.4a UWB Systems, in: 2011 Fifth International Conference on Innovative Mobile and Internet Services in Ubiquitous Computing (IMIS), Seoul, Korea, June 30-July 2, 2011.
[9]
[10]
[11]
[12]
[13]
[14]
X.W. He, Y. Xiao, Y.M. Wang, A LSSVR three-dimensional WSN nodes location algorithm based on RSSI, in: 2011 International Conference on Electrical and Control Engineering (ICECE), Yichang, Hubei, China, Sept. 16-18, 2011. O.G. Adewumi, K. Djouani, A.M. Kurien, RSSI based indoor and outdoor distance estimation for localization in WSN, in: 2013 IEEE International Conference on Industrial Technology (ICIT), Cape Town, South Africa, Feb. 25-28, 2013. K. Benkic, M. Malajner, P. Planinsic, Z. Cucej, Using RSSI value for distance estimation in wireless sensor networks based on ZigBee, in: 15th International Conference on Systems, Signals and Image Processing, Bratislava, Slovak Republic, June 25-28, 2008. J. Neto, Y. Yang, I. Glover, Plausibility of practical low-cost location using WSN path-loss law inversion, in: IET-WSN IET International Conference on Wireless Sensor Network, Beijing, China, Nov. 15-17, 2010. C.C. Hsiao, Y.J. Sung, S.Y. Lau, C.H. Chen, F.H. Hsiao, H.H. Chu, P. Huang, Towards long-term mobility tracking in NTU hospitalâ&#x20AC;&#x2122;s elder care center, in: 2011 IEEE International Conference on Pervasive Computing and Communications Workshops (PERCOM Workshops), Seattle, USA, March 21-25, 2011. A. Rice, R. Harle, Evaluating Lateration-based positioning algorithms for fine-grained tracking, in: Proceedings of the 2005 joint workshop on Foundations of mobile computing, New York, USA, 2005. A. Shareef, Y. Zhu, M. Musavi, Localization using neural networks in wireless sensor networks, in: 9th International Symposium on Communications and Information Technology, Icheon, Korea, Sept. 28-30, 2009. J.N. Tian, Z. Xu, RSSI localization algorithm based on RBF neural network, Software Engineering and Service
Performance Evaluation of Lateration, KNN and Artificial Neural Networks Techniques Applied to Real 1521 Indoor and Outdoor Location in WSN Science (ICSESS), in: 2012 IEEE 3rd International Conference on Software Engineering and Service Science (ICSESS), Beijing, China, June 22-24, 2012. [15] M.S. Rahman, P. Youngil, K.D. Kim, Localization of Wireless Sensor Network using artificial neural network, in: 9th International Symposium on Communications and Information Technology, Icheon, Korea, Sept. 28-30, 2009. [16] K. Langendoen, N. Reijers, Distributed localization in wireless sensor networks: A quantitative comparison, The International Journal of Computer and Telecommunications Networking 43 (2003) 499-518. [17] J. Zheng, A. Dehghani, Range-Free Localization in Wireless Sensor Networks with Neural Network Ensembles, Journal of Sensor and Actuator Networks 1
(2012) 254-271. [18] H.K. Maheshwari, A.H. Kemp, Comparative performance analysis of localization using optimal and sub-optimal Lateration in WSNs, in: Third International Conference on Next Generation Mobile Applications, Services and Technologies, Cardiff, Wales, Sept. 15-18, 2009. [19] RStudio [Online], http://www.rstudio.com/ (accessed December 19, 2013). [20] TinyOs Home Page, http://www.tinyos.net/ (accessed February 09, 2013). [21] R.Q. Min, D.A. Stanley, Z.N. Yuan, A. Bonner, Z.L. Zhang, A Deep Non-linear Feature Mapping for Large-Margin KNN Classification, in: Ninth IEEE International Conference on Data Mining, Miami, FL,USA, Dec. 6-9, 2009.
Journal of Communication and Computer 10 (2013) 1522-1528
Performance Evaluation of Quicksort with GPU Dynamic Parallelism for Gene-Expression Quantile Normalization Roberto Pinto Souto1, Carla Osthoff1, Douglas Augusto1, Oswaldo Trelles2 and Ana Tereza Ribeiro de Vasconcelos2 1. National Laboratory for Scientific Computing Petropolis, Petropolis 25651-075, Brazil 2. Computer Architecture Department, University of Malaga, Malaga 29017, Spain
Received: November 09, 2013 / Accepted: December 10, 2013 / Published: December 31, 2013. Abstract: High-density oligonucleotide microarrays allow several millions of genetic markers in a single experiment to be observed. Current bioinformatics tools for gene-expression quantile data normalization are unable to process such huge data sets. In parallel with this reality, the huge volume of molecular data produced by current high-throughput technologies in modern molecular biology has increased at a similar pace, challenging our capacity to process and understand data. On the other hand, the arrival of CUDA (compute unified device architecture) has unveiled the extraordinary power of GPUs (graphics processing units) to accelerate data intensive general purpose computing more and more as time goes by. In this work, we have evaluated the use of dynamic parallelism for ordering gene-expression data, where the management of kernels launching can be done not only by the host, but also by the device. Each sample has more than 6.5 million genes. We optimized the Quicksort parallel implementation available in the CUDA-5.5 Toolkit Samples and compared the performance of the sequential Quicksort algorithm from the GNU C Library (glibc) and with the parallel radix sort implementation available in the CUDPP-2.1 library. The Quicksort parallel implementation is designed to run on the GPU Kepler architecture, which supports dynamic parallelism. The results show that in the studied application the GPU parallel version with dynamic parallelism attains speed-ups in the data-sorting step. However, to achieve an effective overall speed-up considering the radix sort algorithm, performance of the whole application needs further optimizations. Key words: Bioinformatics, GPU, sort algorithms, dynamic parallelism, data normalization.
1. Introductionď&#x20AC; The huge volumes of molecular data produced by current high-throughput technologies in modern molecular biology pose challenging problems in our capacity to process and understand data. On one hand, current bioinformatics tools for gene-expression quantile data normalization are unable to process such huge data sets [1]. On the other hand, the arrival of NVIDIA CUDA (compute unified device architecture) has unveiled the extraordinary power of GPUs (graphics processing units) to accelerate data-intensive general purpose computing more and more as time Corresponding author: Roberto Pinto Souto, D.Sc., research fields: high performance computing, parallel programming and GPGPU. E-mail: rpsouto@lncc.br.
goes by. This work contributes to the field of large-scale processing of molecular data by investigating the application of GPU dynamic parallelism to a biomedical application: the gene-expression quantile normalization for high-density oligonucleotide array data based on variance and bias, the Q-norm algorithm [2]. When running genetic experiments that involve multiple high-density oligonucleotide arrays, it becomes crucial to remove sources of variation between samples of non-biological origin. Normalization is a process for reducing this variation, with the aim to make different samples comparable. In this work, we implement features from the recent CUDA-5.5 technology to further increase the high performance
Performance Evaluation of Quicksort with GPU dynamic Parallelism for Gene-Expression Quantile Normalization
implementation of the Q-norm algorithm. A previous work [3] reveals advantages and drawbacks of using the GPU as a target compute device, providing lessons that benefit a broad set of existing genetic applications that either are based on those pillars or have similarities with their procedures. In a nutshell, our work reveals that dynamic parallelism provides GPU scalability on the recursive Quicksort algorithm. However, it is not as efficient as the most recent radix sort implementation [4], which is state-of-the-art sort algorithm for GPU. The rest of the paper is organized as follows. Section 2 describes the Q-norm algorithm and its input data set. Section 3 outlines the CUDA programming model and Section 4 explains the parallelization of Q-norm on GPUs. Section 5 presents related works. Section 6 discusses the performance numbers obtained when running our parallel Q-norm implementation. Finally, Section 7 provides conclusions of this work as well as some insights for future work.
2. Q-norm: An Algorithm for Normalizing Oligonucleotides The ultimate goal of the Q-norm method is to make the distribution of probe intensities for each array in a set of samples the same, under the assumption that there is an underlying common distribution for all those samples. This suggests that one could give two disparate data sets the same distribution by transforming the quantiles of each data set to have the same value, which will be their average value. Extending this idea to N dimensions gives us a method of finding a common distribution from multiple data vectors. Let qk = (qk1, …, qk1) for k = 1, …, nG be the vector of the k-th quantiles for all nE array experiments/samples of length nG that compose matrix X of dimensions nG × nE, where each sample is a column. The quantile normalization goes as follows: (1) Sort each column of X to give Xsort;
1523
(2) Take the means across rows of Xsort and assign this mean to each element in the row to get X´sort; (3) Produce Xnorm as output by rearranging each column of X´sort to have the same ordering as original X. This method forces the values of quantiles to be equal, which may cause problems in the tails where it is possible for a probe to have the same value across all the array samples. However, this case is unrealistic since probe set expression measures are typically computed using the value of multiple probes.
3. General Purpose Graphics Processing Unit with CUDA Modern GPUs are powerful computing platforms that reside at the extreme end of the design spaces of throughput-oriented architectures, allowing hardware scheduling of a parallel computation to be practical. Their increased core counts and hardware multithreading are two appealing issues that CPUs are quickly adopting, but until both models converge, if ever, we face a transition period of heterogeneous computing where a PC is seen as a bi-processor platform with the GPU acting as an accelerator of data-parallel code shipped from the CPU. This way, the CPU plays the role of a host holding a C program, where I/O is commanded, GPU transfers are controlled and GPU kernels are launched. Such kernels are developed in CUDA, a programming model for general purpose computing on GPUs. 3.1 Programming Model As a programming interface, CUDA consists of a set of C language library functions, and the CUDA specific compiler generates the executable for the GPU from a source code where the following elements meet: (1) A program is decomposed into blocks that run logically in parallel (physically only if there are resources available). Assembled by the developer, a block is a group of threads that is mapped to a single
1524
Performance Evaluation of Quicksort with GPU dynamic Parallelism for Gene-Expression Quantile Normalization
multiprocessor, where they access a private shared memory. (2) All threads of concurrent blocks on a single multiprocessor divide the resources available equally amongst themselves. The data are also divided amongst all of the threads in a SIMD fashion with a decomposition explicitly managed by the developer. (3) A kernel is the code to be executed by each thread. Conditional execution of different operations can be achieved based on a unique thread ID. In the CUDA model, all of the threads can access all of the GPU global memory, but, as expected, there is a performance boost when threads access data resident in shared memory, which is explicitly managed. In order to make the most efficient usage of the GPUs computational resources, large data structures are stored in global memory and the shared memory should be prioritized for storing strategic, often-used data structures.
4. Recursive Quicksort Algorithm and Dynamic Parallelism Dynamic Parallelism allows the GPU to operate more autonomously from the CPU by generating new work for itself at run-time, from inside a kernel. The concept is simple, but the impact is powerful: It can make GPU programming easier, particularly for algorithms traditionally considered difficult for GPUs, such as divide-and-conquer problems. With Dynamic Parallelism, the GPU is able to generate new work for itself without involving the CPU at all. This permits dynamic run-time decide what to do next, enabling more complex algorithms possible than previously, while simultaneously releasing the CPU to conserve power or perform other work. To handle this dynamic work, NVIDIA created a new hardware technology in Tesla K20 GPUs called the GMU (grid management unit). This manages the complexities of dynamic execution at hardware speed
launching, suspending and resuming kernels, as well as tracking dependencies from multiple sources. A layer of system software running on the GPU interacts with the GMU, enabling the CUDA Runtime API (application programming interface) to be used from within a kernel program. The goal of Quicksort is to sort an array of numbers, and it begins by picking a “pivot” value to be used to partition the array into two smaller arrays: one with values less than the pivot, and one with values equal or greater. After partitioning the initial array, the algorithm then launches two new quick sorts on the two new arrays, producing four sub-arrays and so on until each sub-array contains just a single value; the result is put together and the algorithm terminates. It is a classic “divide-and-conquer” algorithm in which it breaks the problem into even smaller pieces and solves them recursively. In Quicksort, the information needed to sort each stage depends on the stage before it. Without Dynamic Parallelism all of the launches must take place from the CPU, which means that the details of what to launch next must be passed back to the host after each stage. With Dynamic Parallelism the GPU performs its own launches on-the-fly, enabling each Quicksort to launch its two sub-sorts as soon as it has finished. There are no complex overheads like the CPU/GPU stack exchange, and no need for all the host code which manages the launches.
5. Related work Typically, the Quicksort algorithm used does not perform well on graphics processors. With the emergence of CUDA, Cederman and Tsigas [5] succeeded in developing an efficient and well suited to GPU implementation. This becoming the fastest ever developed so that moment for graphics processors. It was superseded by the radix sort for GPU done by Satish et al. [6]. It fully exposes fine-grained parallelism and decomposes the computation into independent tasks that perform minimal global
Performance Evaluation of Quicksort with GPU dynamic Parallelism for Gene-Expression Quantile Normalization
communication, exploiting the high-speed on-chip shared memory. Nowadays, the fastest sort algorithm for the GPU is the radix sort implementation developed by Merril and Grimshall [4]. Their algorithm uses a very efficient strategy that employs prefix scan locally by the kernel, multiple concurrent prefix scans and low rate of synchronization and communication within algorithm phases. In the field of gene-expression experiments, ultra high-density micro-arrays have jumped from 50.000 probes contained in a simple array to more than 5 million genetic markers. Not only have the number of data points per sample increased but also the number of samples per experiment grows at a rapid rate, because of reductions in data acquisition costs. The reference software for gene-expression data normalization, R-Bioconductor [7, 8], is unable to
1525
GPU generations, it was possible only to do that by device kernel functions (i.e. the type of kernel that is launched only by another device kernel). In both CPU and GPU implementations, the Quicksort algorithm receives as input to be sorted an array of struct containing information of the value and original index of each element of the data set, as shown in Fig. 1. It was done this way so that it would be possible to recover the original position of the data after obtaining the average of all sorted experiments. As shown in Fig. 2, the Quicksort function is the interface to the glibc library function qsort. The input parameters are the data type struct array (input), the number of array elements (nG), the size in bytes of the array (sizeof(dstru)), and the comparison function (cmp), which returns the difference between the values of the field val.. 6.1 The Input Data Set
manage these large data sets and sequential computing fails to handle the analysis of high-throughput transcriptomic data. Rodriguez et al. [3] discuss the pros and cons when targeting the GPU and provide mechanisms for effective task scheduling, data partitioning and GPU mapping.
6. Results This
section
presents
the
computational
The high computational cost and memory requirements of the Q-norm method result from its application to huge data sources, which in our case means that arrays composed of more than 6 million gene-expression values. Each array contains positive integers values that were obtained by high-density oligonucleotide microarray technology which provided by the Affymetrix GeneChip infrastructure [9], which is widely used in many areas of biomedical research.
performance of the Q-norm method using the Quicksort algorithm from the GNU C Library (glibc)—qsort function as defined in stdlib.h—and also the GPU implementation adapted from the NVIDIA CUDA-5.5 Toolkit Samples. The main point of this GPU implementation of Quicksort is the use of dynamic parallelism, which is a new feature brought by the NVIDIA’s Kepler-based Tesla K20 GPU. This feature allows a global kernel function (i.e. the type of kernel that is launched by the host) to make calls to itself recursively, analogous to how it is done by the standard recursive Quicksort implementation on the CPU. Previously in earlier
Fig. 1 Struct type that carries the value and the index of sorted data.
Fig. 2 Struct type used in the stdlib.h qsort and the interface function Quicksort, which calls qsort.
1526
Performance Evaluation of Quicksort with GPU dynamic Parallelism for Gene-Expression Quantile Normalization
Those integers are the target numbers to normalize, usually by means of some kind of averaging over every array element placed in the same quantile [9, 11]. In the experimental case of this work, it is considered that nE = 32 samples (number of arrays), with each one composed of nG = 6,553,600 gene-expression values. The input data set was taken from the GEO (gene-expression omnibus) Web repository [1] as submitted by Affymetrix under the GeneChip Human Mapping 500K Array Set (platform GPL3718). 6.2 Performance Analysis The configuration of computing resources is described in Table 1. The CPU is an Intel Xeon E5650 and the GPU is an NVIDIA Tesla K20c. The experiments were carried out by first using exclusively the CPU in serial mode (a single core) for all stages of Q-norm. Then, the second experiment used a mixed approach in which only the demanding sorting stage computed by Quicksort was executed in parallel on the GPU (the other stages remained on the CPU, in sequential). The third and last experiment also employed a mixed approach in which the sorting stage was executed in parallel, by a state-of-the-art radix sort algorithm from the CUDPP-2.1 library, presented in Table 2. The performance of the serial computational method Q-norm, i.e. running on a single CPU core, is presented in Table 3. Observe that, in fact, the main hotspot of Q-norm is the step of sorting the data held by the quicksort routine approximately 87% of the runtime which is due to this task. When replacing the serial quicksort function with a corresponding parallel Quicksort version from the CUDA-5.5 Toolkit Samples, which is optimized for the Kepler architecture through dynamic parallelism, there was a decrease in the runtime. This work further decreases Quicksort runtime compared with our previous work [12], where employed 16 recursive was
levels, by increasing this number of levels up to 22. In Table 3, the sorting algorithm line presents the CPU sequential version, the GPU optimized Quicksort algorithm and the GPU radix sort execution time. The speed-up achieved by the Quicksort routine alone was about 14 times faster than the exclusively serial code. According to Table 4, the speed-up achieved by the optimized Quicksort algorithm increases up to 22.7 times. However, when we also consider the cost of data transfer between the CPU and the GPU, this gain drops to about 14 times faster (Table 5). Moreover, when considering the total execution time of the method Q-norm, the speed-up is further reduced to about 4.8 times (Table 6) Although relatively low, the overall gain of performance is slightly larger than half the maximum possible speed-up for this application. According to Amdahlâ&#x20AC;&#x2122;s law, the speed-up of a program is limited by Table 1
Computational resources.
Processor type Model Cores Clock speed SP flop/clock Rpeak
CPU Intel Xeon E5650 6 2.66 GHz 26 511 GFlop/s
GPU NVIDIA Tesla K20c 2496 0.71 GHz 2 3544 GFlop/s
Table 2 Sort algorithms used. Algorithm QsortCPU QsortGPU Radixsort
Function qsort quiksort_cdp cudppRadixSort
Speed-up CNU glibc library CUDA-5.5 samples CUDPP-2.1 library
Table 3 Q-Norm performance profile of CPU and GPU sorting algorithms. Time in seconds CPU Main Q-norm main LoadFile sortingAlgorithm AccumulateRow BackPos StoreFile CudaMemcpy
QsortCPU 33.69 33.69 0.26 29.51 0.31 2.06 0.49 -
GPU QsortGPU 6.96 6.80 0.38 1.30 0.25 2.45 0.45 0.80
Radixsort 5.95 5.78 0.29 0.23 0.22 1.79 0.42 0.89
Performance Evaluation of Quicksort with GPU dynamic Parallelism for Gene-Expression Quantile Normalization Table 4 Speed-up among CPU and GPU sorting algorithms: Only kernel function time. SortingAlgorithm QsortCPU QsortGPU RadixSort
Time (s) 29.51 1.30 0.23
Speed-up 1.0 22.7 128.3
Table 5
Speed-up among CPU and GPU sorting algorithms: Kernel function + data transfer time. SortingAlgorithm QsortCPU QsortGPU RadixSort
Time (s) 29.51 2.09 1.12
Speed-up 1.0 14.1 26.3
Table 6 Speed-up among CPU and GPU sorting algorithms: program main function time. SortingAlgorithm QsortCPU QsortGPU RadixSort
Time (s) 33.69 6.96 5.95
Speed-up 1.0 4.8 5.7
the portion not parallelized, and this limit is given by the inverse of this portion. Since the parallelized code portion (serial quicksort routine) corresponds to 87.6% of the overall execution time, the speed-up is limited by the inverse of 0.124 (= 1.000 – 0.876). Thus, the speed-up will not exceed 8 (= |8.06|) When replacing the parallel quicksort function by a corresponding parallel radix sort version from the CUDPP-2.1 library, speed-up further increases from 4.8 to 5.7 (Table 6).
7. Conclusions This work has presented methods for computing a quantile normalization of high-density oligonucleotide array data on GPUs. Our approach focuses on CUDA-5.5, which allows for exploiting dynamic parallelism, and also takes advantage of the expressive processing power offered by the GPU Kepler architecture. Current solutions of Q-norm based on the R-Bioconductor software fail to process huge data volumes, so we believe that our contribution represents a step forward to provide computational support to a generic Q-norm for large microarray data
1527
sets as well as a low-cost and high-performance alternative to high-end workstation systems. We optimized Quicksort dynamic parallelism algorithm and compared it with the state of the art radix sort algorithm. Although the optimized Quicksort dynamic parallelism algorithm presents impressive Quicksort serial performance, the improvements are still not competitive compared with the radix sort algorithm. These experiments succeeded with increases from 16 up to 22 recursive levels with the Quicksort algorithm. For future work, we plan to use recent Kepler K40 technology in order to improve the recursive levels from the Quicksort algorithm and hopefully achieve a competitive performance compared with radix sort algorithm.
Acknowledgments The authors would like to thank the Brazilian National Research Founding Agency—CNPq (301877/2013-0), Rio de Janeiro State Research Founding Agency—FAPERJ (E-26/102.025/2009) and CUDA Teaching Center Program. This work has been partially supported by the EU-FP7 IAPP Project Mr. SymBioMath (324554) and the Brazilian “Science without Borders Programme” (401981/2012-6).
References [1] [2]
[3]
[4]
The GeneChip Human Mapping 500K Array data set Submitted to GEO by Affymetrix. B.M. Bolstad, R.A. Irizarry, M. Astrand, T.P. Speed, A comparison of normalization methods for high density oligonucleotide array data based on variance and bias, Bioinformatics 19 (2003) 185-193. A. Rodriguez, O. Trelles, M. Ujaldon. Using graphics processors for high performance normalization of gene-expression, in: 2011 IEEE 13th International Conference on High Performance Computing and Communications (HPCC), Banff, AB, Sept. 2-4, 2011, pp. 599-604. D. Merrill, A. Grimshaw, High performance and scalable radix sorting: A case study of implementing dynamic parallelism for GPU computing, Parallel Processing Letters 21 (2011) 245-272.
1528 [5]
[6]
[7]
[8]
[9]
Performance Evaluation of Quicksort with GPU dynamic Parallelism for Gene-Expression Quantile Normalization D. Cederman, P. Tsigas, GPU-Quicksort: A practical Quicksort algorithm for graphics processors, Journal of Experimental Algorithmics 14 (2009) 1-22. N. Satish, M. Harris, M. Garland, Designing efficient sorting algorithms for manycore GPUs, in: IEEE International Symposium on Parallel & Distributed Processing, Rome, Italy, May. 23-29, 2009, pp.1-10. R. Ihaka and R. Gentleman, R: A language for data analysis and graphics, Journal of Computational and Graphical Statistics 5 (1996) 299-314. R. Gentleman, V. Carey, D. Bates, et al., Bioconductor: open software development for computational biology and bioinformatics, Genome Biology 5 (2004) R80. J.A. Warrington, S. Dee, M. Trulson, Large-Scale Genomic Analysis Using Affymetrix GeneChip,
Microarray Biochip Technologies, BioTechniques Books, New York, USA, 2000, pp. 119-148 (Chapter. 6). [10] Statistical Algorithms Referende Guide, Technical report, Affymetrix, 2001. [11] R.A. Irizarry, B. Hobbs, F. Colin, Y.D. Beazer-Barclay, K. Antonellis, U. Scherf, et al., Exploration, normalization and summaries of high density oligonucleotide array probe level data, Biostatistics 4 (2003) 249-264. [12] R.P. Souto, C. Osthoff, A.T. Vasconcelos, D.A. Augusto, P.L. Silva Dias, A. Rodriguez, et al., Applying GPU dynamic parallelism to high-performance normalization of gene-expressions, in: 4th Workshop on Applications for Multi-Core Architectures (WAMCA 2013), Porto de Galinhas, Brazil, 2013.
Journal of Communication and Computer 10 (2013) 1529-1553
A Privacy Taxonomy for the Management of Ubiquitous Environments Valderi Reis Quietinho Leithardt1, 2, Guilherme Antonio. Borges2, Anubis Graciela de Morais Rossetto2, Carlos Oberdan Rolim2, Claudio Fernando Resin Geyer2, Luiz Henrique Andrade Correia3, David Nunes4 and Jorge Sa Silva4 1. Institute of Technology, National Service of Industrial Training (SENAI), Porto Alegre 91140-000, Brazil 2. Institute of Informatics, Group of Parallel and Distributed Processing (GPPD), Federal University of Rio Grande do Sul (UFRGS), Porto Alegre 91509-900, Brazil 3. Federal University of Lavras, Lavras 37200-000, Brazil 4. Coimbra University, Coimbra 3030-290, Portugal Received: October 10, 2013 / Accepted: November 15, 2013 / Published: December 31, 2013. Abstract: Pervasive and ubiquitous environments must handle the detection and management of users, devices and services, while guaranteeing the privacy of both the users and the environment itself. Current techniques for handling privacy found in the literature treating the subject in various ways, while concentrating on the device management, communication protocols, user profiles and environmental access. This paper examines a control model for privacy in pervasive environments from the perspective of the environment. A prototype was devised and tested to validate the generic model of privacy which was also used to compare taxonomic concepts in the literature. Moreover, the prototype was devised and tested to validate the generic model of privacy for control and manage various users, devices and environments and so on. The prototype was based on Percontrol (a system for pervasive user management), which was only intended to identify users using Wi-Fi, and now it is capable of managing temperature, luminosity and other preferences, measured by a WSN (wireless sensor network) embedded to Percontrol, and the data treatment is done by an ANN (artificial neural network). Results confirmed the viability of device detection with Wi-Fi, Bluetooth and RFID (radio frequency identification) for an increases slight of the latency in registering new devices on the system. Key words: Percontrol, generic privacy model, ubiquitous environment, artificial neural network.
1. Introduction In a pervasive environment, computational resources are come omnipresent in people’s daily lives and are all interconnected with the objective of providing accurate information and services, regardless of the time or place [1]. Thus, the environment is filled with computational devices that are ingrained in such a way that using them becomes “second-nature”, and thus creates the illusion that they just “disappear” and become a normal part of people’s Corresponding author: Valderi Reis Quietinho Leithardt, Ph.D. candidate, research fields: privacy and ubiquitous computing. E-mail: valderi.quietinho@inf.ufrgs.br.
daily lives. In recent years, a new computational paradigm is emerging with the emergence of mobile devices, for which the calculation is highly dynamic and must adapt fast to environmental changes. This phenomenon is caused by the user’s own mobility and in situations where the processing power exists within small multi-function mobile devices such as mobile phones, smartphones and PDAs (personal digital assistants) [2]. According to Ref. [3], ubiquitous computation is the kind of computation that makes life simpler; digital environments can sense and process information, by being adaptable and becoming
1530
A Privacy Taxonomy for the Management of Ubiquitous Environments
pro-active towards human needs. Weiser [1] forecast new systems and environments that would be full of computational resources capable of providing services and information whenever necessary (“everywhere, every time computing”). Thus, he proposed a continuous integration between the environment and technology, with the aim of helping people carry out their everyday activities within this environment [1]. The current trend in computing is the usage of “invisible” computers, where the man-machine interaction is governed by non-traditional means; instead of the traditional keyboard and mouse, touchscreen and motion controllers are quickly becoming the standard input mechanisms. Computers are now set up with the aim of responding to user-stimulus, without the need for direct user interaction. This concept is close to the idea of Pervasive computing, since machines are distributed within the environment in a non-perceptible fashion; through the use of sensors or other means of communication such as RFID [4], Bluetooth or WSNs (wireless sensor networks) [5]. These machines communicate between themselves, the users and the environment, and their tasks are modeled to suit everyone’s needs in a better way. Pervasive computation should be context-aware and intelligently adapted to finding better solutions to the most diverse situations. Pervasive computation deals with many situations that have no equivalent in traditional computation: Common among these are changes in the presence of users, location conditions, service availability (such as weather forecasts or clock synchronization), and computational context. Privacy plays an important role in these situations, since a user might not want to be located or share his data during a certain time. These requirements should be met by the pervasive environment, by reducing the processing of unnecessary data and increasing the overall security level and the performance of service management
tasks. We propose a privacy control model for pervasive/ubiquitous environments to properly address the requirements of this pervasive computation; this handles as many requirements related to the environment as possible. In the current literature, several scientific papers were found which adopted an approach to privacy control that involved using different techniques that focused on the user himself or the devices, communication services and privileges which he is able to access. However, we consider a broad range of different scenarios present in the real world (e.g. churches, libraries and football stadiums), although no one specific work addresses all the necessary requirements that can be found in all the existing scenarios. The rules and regulations that govern each one of us, are mostly determined by the context of our situation, and the context is closely tied to the environment where we currently reside. Thus, most privacy control mechanisms that focus on the users, devices, communications or services usually lose their validity when removed from the environment for which they were designed. The main contribution made by this work is to outline a novel model for privacy that is focused on the pervasive/ubiquitous environment, and seeks to bring the concept of Pervasive computing closer to the real world. However, solutions for the problem of security in pervasive/ubiquitous computation will not be addressed, such as techniques for avoiding attacks or encryption algorithms; nor will solutions for restricting users and/or devices, or the services and communication mechanisms available to them. This paper is structured as follows. In Section 2, an analysis of the state of the art in the literature is conducted, and some key concepts of privacy in pervasive and ubiquitous environments are defined and discussed. A taxonomy for privacy in ubiquitous environments is outlined in the Section 2.1. This taxonomy is defined and compared with the models
A Privacy Taxonomy for the Management of Ubiquitous Environments
found in the current literature, as show the Section 2.2. In Section 3, the criteria and definitions of the generic model of privacy control proposed for pervasive and ubiquitous environments are analyzed in depth. Section 4 presents a scenario outlined of the application based on the characteristics and definitions listed in the taxonomy and model of privacy. In Section 5, a test-bed was employed to validate the architecture of the proposed application for pervasive and ubiquitous environments. Finally, the conclusions and suggestions for future work are described in Section 6.
2. State of the Art In pervasive environments, there are several problems and challenges that have to be faced, among which, the control and management of privacy stand out. There are several different concepts and definitions of privacy in Ref. [6]. We can cite the ideas introduced by Refs. [7, 8], where privacy is considered to be an abstract and subjective concept that is closely bound up with each individual’s perception of what it represents. According to Ref. [6], privacy can be related to providing protection from threats to one’s personal property, or physical and moral integrity; these needs are not uniform and are influenced by cultural factors such as religion, tradition, customs, education, and the political environment, as well as more personal factors like age, health, occupation and humor, among others. Despite the extensive literature in the area, many questions are still left unanswered, while others still require a great effort to integrate several concepts and techniques into a single. A solution can handle privacy in complex environments. It can be readily appreciated that it is not possible to address every aspect of so many situations, and that this invalidates any definition of a specific privacy context [7]. A context is characterized by data that overlaps both the physical and virtual worlds. People do not
1531
usually regard physical environments (the office, shop floor or stadium) and virtual environments (the computer desktop or mobile phone menu) as separate entities, since objects and processes can be represented in both worlds. Hence, it is necessary to project structures that are capable of representing elements from both the real and the virtual domain. These elements should be represented in a way that is as generic as possible, to assist the creation of environments that can provide better support for associated physical and virtual tasks. This can only be achieved by putting forward taxonomic definitions that allow the isolation of specific parameters and requirements associated with pervasive privacy, a task that will be discussed in the next sub-section. 2.1 Taxonomy of Privacy in Pervasive Environments In this section, we outline a taxonomy for privacy in pervasive environments, as shown in Fig. 1. This is based on the current literature and extends the concepts of privacy by taking into account the context of the environment. In Ref. [9], a few important requirements were set out, in which the pervasive user is described as: Collaborative: The user should provide access to information and services (such as music, videos, personal data, and location) in a collaborative fashion, in order to enhance both his own experience and that of other users, as well as to improve the system in general. Flexible: Users can adjust the level of collaboration to suit the safety levels required by a certain service request. Thus, as a result, there is flexibility between the users and sources of information. Visible: The user provides access to his profile and identity, which may hold several classifications as described in Ref. [9], where the user’s identity may be weak (with a minimum degree of trust), average (medium level of trust) and strong (high level of trust).
1532
Fig. 1
A Privacy Taxonomy for the Management of Ubiquitous Environments
Taxonomy of privacy in ubiquitous environments.
Other characteristics should also be considered, including anonymity; there are several situations where the user’s location cannot or should not be divulged, owing to the nature of his/her occupation or for personal reasons. Controllable and the sharing of data should be controlled by the user. His opinions, characteristics and personal data may change at any time, depending on his everyday decisions and lifestyle. Anonymous: Even if user management modules have access to many types of information about users, the user should always be in a position to decide if he/she does not wish to provide access to certain resources or services to other users anymore, by changing his profile or context. The work carried out by Ref. [10] shows a control mechanism applied to a music sharing service based on the location of Wi-Fi hotspots. This paper adopts a different approach from other privacy studies, since it envisages the possibility of defining types and sizes that can be transferred, depending on the current Wi-Fi spot and location. However, this approach does not handle the environment itself but is only concerned with the Wi-Fi spots within it. In Ref. [11], there is a different solution based on an algorithm that computes an area that depends on the level of data protection required. The use-case was
a hospital environment where user data and location could be shared. However, the study did not predict interactions with the pervasive environment, but only took account of its location [12]. The study also adopted different approaches that could restrict access to certain information, such as document validation. With regard to desirable characteristics for handling privacy in devices, some works investigated user locations by means of GPS, wireless access points or cellular antennas, and made use of coordinates between locations to control access to services within the pervasive environment. The research conducted in Ref. [13] defines the following taxonomic requirements for devices: Accountable: These devices must be registered in the places they frequent regularly, since this helps to reduce the amount of unnecessary data traffic and helps speed up the device identification system. An actual application of this would be the storage of device characteristics that visit a certain environment, such as login credentials or available services. Localizable: Direct access to a data base containing information about users and their devices is necessary to validate basic information. Treatment of other types of information should be handled through the access point that the user is currently
A Privacy Taxonomy for the Management of Ubiquitous Environments
using, in order to ensure security and reliability. Heterogeneous: A single device may use different kinds of communication protocols and provide different kinds of services. Computationally limited: This is a very important factor when dealing with battery-powered or mobile hardware, and it should be taken into account when designing effective privacy mechanisms [14]. Low power processing or limited storage should not be regarded as limiting factors but as challenges to be overcome to achieve computational balance. Energy aware: Depending on the hardware in question, we need to consider the energy consumption used by the application layer, service layer and communication. A pervasive system should always take into account how applications or mechanisms can help reduce the amount of energy consumption. The architecture of the system should always focus on external information processing, leaving the sensing and transmission tasks to more limited devices. The taxonomy definitions for applications and services are based on the work described in Ref. [7], which outlined the desirable requirements for privacy services: Flexible: Users should be capable of defining their own privacy preferences, with different levels of detail for different groups of people. Different kinds of users may have different needs while different groups of people may wish to share information in distinct ways (groups united by their religion do not share the same kind of information as groups united by an interest in sports). Notifiable: The users may want to be notified of, or to scan, any attempts to gain unauthorized access to their contextual information. Hence, it should be possible for the user to create custom notifications for different situations. While in certain scenarios, it may be useful to be notified of every access attempt, other scenarios may require these warnings to be automatically discarded (for example during sleep times, which are not the same for every user and may
1533
vary every day). Controllable functionality: In addition to the access control options (“grant” and “deny”), a third option (“not available”) should also be made available. This option allows users to deny access without the requester being aware of it. This technique is also known as “plausible deniability”. Controllable accuracy: Users may adjust the temporal and spatial precision of their context information. This usually applies to the user’s mobility, availability and daily tasks, where information is constantly being changed or updated. Controllable access: The users should be able to block access to any contextual information at any time. As a basic security routine, the system itself should be able to issue a warning that the user or his device are under a security threat and block the sharing of any kind of information. The user may also find himself in unknown public places where information sharing is not recommended. According to Ref. [15], traditional autonomous computing and small networks depend on user authentication and access control to guarantee security. These interaction-dependent methods have certain rules and regulations that restrict the ability to access, use, modify or visualize resources. However, mobile users need to be able to access hosted resources and services at any time from any place, which leads to serious security risks and access control problems. With these challenges in mind, Ref. [12] proposes a solution based on the management of trust which involves the adoption of security policies and the assignment of credentials to external entities. These credentials are checked to see if they conform to the defined policy, while the level of trust in each of these entities is validated through third-party input (the feedback from other entities, for example). The level of trust attributed to an entity can be correlated with the level of access they are given. Despite being valid, the proposed solution does not handle all the necessary requirements shown in Fig. 1 of its
1534
A Privacy Taxonomy for the Management of Ubiquitous Environments
taxonomy, or does it focus on the pervasive environment. Some other characteristics associated with pervasive Services and applications are shown in Ref. [16]: Emergency overridable: The users should be able to define the exceptional policies that precede any other kind of privacy policy. Just as in the real world where we face factors that are beyond our own control, it is necessary to define rules that must be complied with in priority situations, such as emergency phone calls between family members at inopportune times or from call-restricted locations. Simple: Another point is that the users should not be bothered with too many configuration options for their privacy preferences. Basic usability guidelines suggest that no one wants to navigate through a lot of interfaces and menus to configure a particular functionality. Thus, the system should be able to store relevant configuration information, for example the user’s most accessed functionalities, which can then be used to generate useful “shortcuts”. Efficient: The handling of privacy concerns should not cause a significant delay in communication or a heavy processing load for the context providing services; Discoverable: The application should meet the necessary requirements and provide the parameters for the discovery and offer of available services to the users. The discovery of services and the mechanism to make them available should be both omnipresent and automatic, in the sense that it should not be necessary to reconfigure the parameters of a device for each new situation. Available: The application should have control over the usage of services, to ensure that all users, devices, communications and services in the pervasive environment can enjoy equal access to a greater amount of information. The work carried out by Ref. [16] shows that several problems were encountered when dealing with
the management of security in pervasive applications and services. However, we found that its taxonomic description of communication, based on protocols and services, represents its most distinguishing feature. The taxonomy to communication is described as: Varied: The communication should adapt to as many distinct devices and environments as possible, and should be able to exchange data by means of different communication media without the need for user intervention. Adaptable: The communication should control which protocols can be used in each environment, to reduce the risk of message losses and processing requirements. Confidentiality: The infrastructure must be capable of handling performance, certification control, login mechanisms and other management functions so that it can provide pervasive communication in a secure and reliable way, as described in Ref. [17]. Scalable: A communication protocol should simultaneously serve as many users and devices as possible and provide services in many different environments while maintaining a satisfactory level of quality. Automated: The communication should be able to support mobility and remain adaptive by using unicast, broadcast or multicast communication channels, without the need for user intervention. Even in the area of taxonomic descriptions for pervasive computation, the work in Ref. [18] establishes a framework and middleware architecture for pervasive computation. This study argues that a fundamental demand for pervasive computing requires the automatic physical integration of hardware devices. However, this work treats the infrastructural requirements of pervasive software in a general sense, and does not specifically address the question of pervasive environments from a computational standpoint. The work in Ref. [19] explores different forms of communication and distinct infrastructures that
A Privacy Taxonomy for the Management of Ubiquitous Environments
support several requirements and characteristics for pervasive computation: scalability, heterogeneous environments, integration, contextual invisibility, awareness and contextual management, which have been described as the main challenges that had to be addressed by pervasive computation. The work described in Ref. [20] addresses the modeling of systems in the area of pervasive computation. It contains a study on the issue of privacy, which is used as a means to extend previous research work that devised a meta-model to be used as a basis for the construction of ubiquitous systems. The extension proposed in Ref. [20] seeks to specify privacy at user-level for ubiquitous environments. Although, it can be claimed that this study has made a considerable contribution to the state-of-the-art, the proposed approach does not directly deal with privacy in the ubiquitous environment, but rather, is concerned with the ubiquitous user within the environment. Moreover, we lay down a few basic requirements for pervasive environments, which include the following: Context aware: It is one of the most important factors in intelligent environments since it makes a ubiquitous system as minimally invasive as possible. The system and the environment should be able to recognize the user’s current status and adapt their behavior accordingly, as described in Ref. [2]. For example, a user that enters a pervasive space should be automatically identified and have access to the services and environmental configurations that should be available to him. An interesting case of the importance of privacy in pervasive environments is described in Ref. [21], where a British woman found out about her husband’s infidelity through Google Street View, thanks to the customized number plate on the husband’s car. Since its launch, Google Street View has been the target of complaints and was severely criticized for (accidentally) obtaining pictures of people performing acts meant to remain private, without their consent or knowledge. If one thinks of the world of pervasive
1535
computation and sees the car as a pervasive environment, it is possible to configure the car in a way that prevents its location from being published. This could be carried out, for example, by equipping the car with a RFID chip that is read by the Google Street View’s vehicle, and then either gives or denies permission to photograph it. This 2009 case proves that privacy is neither a novel nor a trivial issue. In an attempt to establish a reference model, Ref. [22] provided a taxonomy that was aimed at establishing a new set of QoS metrics for classifying and characterizing WSNs. However, the work did not consider privacy or metrics control. The work in Ref. [23] carries out a review of the state-of-the-art in privacy preservation techniques and a taxonomic analysis of the control of privacy and contextual data for WSNs. Two main categories of privacy-preservation techniques are discussed—data-oriented and context-oriented. This work is particularly useful, since it solves the problem of ensuring privacy for both data in the network and the application context, but, on the other hand, it does not address the challenges associated with the application environment or the pervasive system as a whole, since it is wholly focused on WSNs. We can draw up a new privacy agreement on the basis of this work: Manageable context: The environment should allow the user to share his own data and use the services and information made available by other users and the environment itself. Thus, support for a domain-independent representation of services and information is expected from the environment. With help from the pervasive environment, the user can choose the most suitable and context-appropriate services to achieve his goals. In view of the heterogeneity of possible devices and configurations in ubiquitous computing, the provided services should be accessible from anywhere within the environment and available in as many different formats as possible. According to Ref. [20], an understanding of the
1536
A Privacy Taxonomy for the Management of Ubiquitous Environments
nature and context of human activities is a very important research field in many areas such as psychology, sociology and ergonomics. However, this wide range of involved areas can cause conflicts, since each area offers a different perspective and can propose and explore a different strategy for a better understanding of human activities. An overall understanding of the subject can only be obtained by conducting research in each of these areas; this overview is very important to enable ubiquitous computation to accurately detect, represent and analyze human activity, which is a multidisciplinary challenge of considerable complexity. This reveals the need for combining different systems as a means of providing the ubiquitous environment with information and guidelines on how to act based on human responses. As a result, we can identify a new requirement for obtaining a privacy solution: (1) Interactive: Users must interact with the environment in order to obtain information about it. This interaction should be intuitive, pleasant and adjusted to the environment context. Pervasive computation can lead to a good deal of inconvenience, such as intrusive advertising mechanisms: Most people do not wish to keep being informed about products for sale whenever they pass by a store. One way to circumvent this problem is to only inform registered users that have subscribed to certain types of notifications that match their hobbies and interests, and to allow them to disable and re-enable these notifications at any time. (2) Activity recognition: The user activities can be effectively recognized through specialized activity recognition mechanisms. This information can be used to improve that ability to infer the user’s context from the pervasive environment. Context inference can be used in numerous situations, such as employing automated network mechanisms to suit the user’s needs (e.g. choosing the best wireless interface available, depending on the user’s location and activities).
(3) Registrable: Privacy management in a ubiquitous environment should allow technology to remain very close to individuals and operate in a variety of real scenarios. For the control and registration of environments, it is necessary to have information that describes everything that belongs to the environment, as well as accessibility conditions, available services, shared resources, authorized individuals, devices, communications and applications that allow interaction to occur. However, the rules that govern privacy and control access should always be based on the environment and its definitions. According to the literature, the ubiquity paradigm compels the computation to be invisible, that is, to carry out its operations with the minimum distraction from the task in hand [1-3, 7, 20, 24]. The environment should not have to auto-reconfigure at each login solicitation or change of users. In other words, the environment should be configured in a way that is unique and inherent to its nature and context. If a user goes to a soccer match and sits on the opposite team’s bench, this represents a serious security event for the stadium’s pervasive system, since the user’s location might be hazardous to his well-being. Thus, we need to define the environmental hierarchy of usage: Hierarchical: Pervasive environments should be governed by a set of established rules, much like those in the real world. In general, each environment is managed by a high-ranking entity, which might be the environment’s owner or the person in charge. Companies, churches, schools and even our own homes are governed by a hierarchical tree that determines how individual elements relate to each other in terms of authority. The pervasive environment’s hierarchical taxonomy should follow the same structure as the hierarchies that we find in real world locations. Politically modifiable: With regard to the policies of the environment, it is necessary to enable the user in his current location to have control of input/output
A Privacy Taxonomy for the Management of Ubiquitous Environments
of information, that is, the user’s environment controls the way information and services are shared. For example, a classroom can be configured to receive the teacher as a kind of advanced user and students as simple guests [25]. In the research conducted by Ref. [26], a logical language for expressing security policies called LEPS is set out, which defines a security policies model for access control services, authentication, integrity, privacy, auditing and non-repudiation. The proposal is of value, but it remains focused on users and related groups. However, we can use the concepts introduced by LEPS to create other requirements from the hierarchical model proposed: Intelligent: These mechanisms should have definitions and rules so that the environment itself is enabled to make the decisions intelligently, without human intervention. One of the most popular solutions is the introduction of artificial intelligence mechanisms such as those outlined by Ref. [27], where an inferential information system called MANFIS is described, which allows multiple data input and consequently only one output. In research conducted by Ref. [14], there is a classification taxonomy for different ubiquitous environments, which is supported by two main categories: interactive environments and intelligent environments. The taxonomy classifies all the types of ubiquitous environments which allow interaction with the user in intelligent daily operations. The classification is based on the routine behavior of the user, so the environment has the usual information. The study states that, regardless of the environment, decision-making mechanisms are necessary to maintain control [28]. However, there is a lack of a precise taxonomic definition of what is required of pervasive environments. The authors simply draw on the studies and techniques employed by other researchers to give an idea of what would be an ideal solution. They fail to address the various issues of pervasive environments, and simply conclude that the
1537
pervasive environment should be iterative so that it can meet the pervasive requirements. Portable: In the case of the pervasive/ubiquitous environment, portability is an important requirement that was not addressed by any of the investigated works. It is necessary because restricting an application or service to a single programming language, operating system or other forms of use in pervasive environment, can also be considered to be a kind of sharing and is not an imposition. It is impossible to have pervasive/ubiquitous environments that impose the use of certain technologies, devices or software. Certainly, there are situations in which some regulations are necessary when dealing with the topics discussed above, but in these cases it is necessary to find generic solutions where the user employs as few of the environment resources as possible. The work in Ref. [29] proposes a model, based on theoretical investigations into personal interrelations, that seeks to embrace human privacy and bring it to the pervasive world. A state of the art that is enhanced by works on human interactions, is used to derive the peoples’ preferences from data on groups of users and the collaborations between them. The model proposed by Ref. [29] is based on user registration and control, where the environment acts in an omnipresent way, and also on the stored information. While the idea is well founded and has several factors that are of value to this work, the model does not address the question of the pervasive environment, but focuses on the people who inhabit it, and is shaped according to their preferences. In Ref. [30], research is conducted into several privacy issues addressed in the context of HCI (human computer interaction) and, based on this research, a number of trends in the area are defined. As its main contribution to this research study, the work addresses several questions, including the protection of the pervasive environment. We examine this item (and how it can be adapted) from the perspective of privacy
1538
A Privacy Taxonomy for the Management of Ubiquitous Environments
control in the environment, since the focus of Ref. [30] is not on the privacy of pervasive environments. The work in Ref. [31] offers a solution based on authentication that takes account of various technological scenarios, such as RFID use, while suggesting a single mechanism to manage different authentication protocols in ubiquitous environments. However, these mechanisms are only concerned with performing authentication iterations over the pervasive system, and taking note of possible changes in the environment. These limitations make it necessary to add additional systems so that the pervasive environment can share information, and this restricts the feasibility of adopting the proposal. There are some other examples that have a bearing on the aims of this study with regard to control of privacy and, hence, the sharing of data, describing the privacy of obtained information and the classifying the captured flows as public or private. Some early work had already expressed concern over what is a recurring problem [2]. 2.2 Comparison Table of the State of the Art Among the studies reviewed, there is a focus on devices as well as their means of communication, and also on the relationship that is necessary for control of privacy. However, these studies do not deal directly with the privacy control of the requirements and relationships within the environment. In view of this, we seek to define two key areas, the handling of environmental characteristics and the sharing of information that will be made available to the user. Devices and users will have to adapt to the environment context and not the other way around. This approach aims to adapt the pervasive context to the real world in which we live in, where a certain environment and its rules are not changed due to the presence of the users within it and their privacy preferences. On the basis of these premises, some works were analyzed, and prominence was given to
those that [12] calculate the userâ&#x20AC;&#x2122;s location by means of the GPS and computes the possible points inside a building, where a particular user may be. It then applies privacy rules to the users, depending on their probable location [12]. While it was an interesting approach at the time, there still remain several drawbacks to this approach. It does not address, for example, the question of the services carried out by the user since it does not know his exact position, nor does it handle the data sharing in larger areas, since the GPS may show some discrepancy between the detected and actual position, in the order of several centimeters. From an application standpoint, this may seem a minor point, but in the case of distinguishing between different divisions, a thin wall that is a few centimeters thick can make the difference between one environment and another. In Ref. [32], service discovery protocols are designed to reduce administrative overhead and increase usability. They can also save pervasive system designers from having to predict and encode all the possible interactions and the states between the devices and applications at design time. By adding a control layer, service discovery protocols seek to simplify the system performance. This work shows good taxonomic definitions of communications and services. However, the proposed solution is focused on control protocols and data generated by the device connected to the user, and does not provide definitions and descriptions aimed at controlling the pervasive environment. A comparison of different research studies is shown in Table 1 to achieve a better overview of the state of the art and the proposed model. The Table shows several of the requirements for pervasive/ubiquitous environments collected from the researched literature. The research explores these requirements extensively and provides taxonomic definitions for each, as shown in Fig. 1. On the basis of this study and the state of the art, we have defined a model for privacy in pervasive/ubiquitous environments, called the GMP (generic model of privacy)
A Privacy Taxonomy for the Management of Ubiquitous Environments
Table 1
1539
Comparison between work-related.
Approaches LPPC SDPCE GPASCRM TBSPCE EPAECU TGSIUC LEPS SLPPC INFOPOINT PRISM MGSPPSC TSPPC ECMPPCE ASCLP FSSD SDPCE TUCE CAUASPB EUPHCI GMP
User Approached Approached Approached Development Development Development Approached Approached Approached Not described Approached Not approached Approached Not approached Development Development Approached Approached Approached Approached
Device Not approached Not approached Not approached Approached Not approached Development Development Approached Approached Not approached Approached Approached Not approached Not described Not approached Not approached Approached Development Approached Approached
Application Approached Approached Approached Approached Approached Approached Approached Approached Development Development Approached Approached Approached Development Approached Approached Development Approached Approached Approached
(Section 3.3), which is compared with the other models shown in the Table 1. We have used a number of definitions, such as: Approached: The work deals with the item; Not Approached: The work does not address the item; Not Described: Information on the item is not found; Development: The item is still under development; this often pointed out in the testing, validation, results or future work sections. Several works describe a particular solution that may be applicable to the pervasive environment, but fail to provide information, testing results or simulations on how to control the pervasive system and all its elements that are found in the environment. In Section 3, we will outline a privacy control model for pervasive/ubiquitous environments.
3. Proposed Model Context awareness, according to Ref. [33], refers to any information that can be used to characterize the
Services Not approached Approached Approached Approached Approached Approached Approached Not approached Approached Development Approached Approached Development Development Approached Approached Approached Approached Not approached Approached
Communication Not approached Development Approached Not approached Approached Development Approached Development Approached Approached Not approached Not approached Approached Approached Approached Approached Approached Approached Approached Approached
Environment Not approached Not approached Not approached Not described Not described Not approached Development Not approached Not approached Not approached Development Not approached Not approached Not approached Not approached Not approached Not described Not approached Not approached Approached
situation of an entity. An entity can be a person, object or place that is considered relevant to the interaction between a user and an application. Also, there are four types of context that are defined by Ref. [8]: (1) Context of Computing (networks and resources); (2) User Context (people, places and objects); (3) Physical Context (lighting, odor, temperature); (4) Temporal Context (hours, days, months). An example of use context is the ability of a device to measure the temperature in a given environment and employ equipment (e.g. air-conditioning) to provide the ideal temperature for the users inside. Another definition is given in Ref. [2], which states that the context of a user in context-aware applications consists of attributes such as physical location, the physiological state, the emotional state, personal history, and daily patterns of behavior, among others, which, if applied to a human assistant, can be used for decision-making without the constant need for the user’s attention. However, there are two serious difficulties related to the development and use of
1540
A Privacy Taxonomy for the Management of Ubiquitous Environments
context-aware applications: the complexity of providing context-aware services and the need to maintain the privacy of contextual information (e.g. the location of the user). These applications generally involve the use of computational contexts (e.g. energy level, bandwidth), personal (e.g., profile, user location) or physical contexts (e.g. temperature, humidity) to provide the customized services that are most appropriate for a particular end-user [7]. In Europe, there are already laws and policies designed to protect the privacy of personal data. For example, the global roaming service for mobile phones encouraged some countries to implement legislative policies aimed at protecting personal privacy. The European Union Directive on data protection [34], which currently comprises the most complete set of privacy laws, has had several updates since the work described in Ref. [7] was published. These updates state that personal information should be: (1) Obtained accurately and within the conditions stipulated by law; (2) Only used for the original purpose specified; (3) Requested in an appropriate manner, that is relevant to the original purpose; i.e. the accuracy of the requested information must not be more specific than what is absolutely necessary to meet the needs of the requester; (4) Kept in a safe place; (5) Accessible to the owner of the information; (6) Destroyed after the purpose of its use has been fulfilled. Some other new policies have been added since then, with an emphasis on public protection. These establish a mandatory legal framework that guarantees the individual right to privacy [34]. This right is ensured through the implementation of measures that must be respected by any organization (including governments and corporations) that deals with personal data during the stages of both the applicationâ&#x20AC;&#x2122;s design and its implementation. These measures cover the processing of personal
data and include provisions relating to the following: (1) security of networks and services; (2) confidentiality of communications; (3) access to stored data; (4) processing of traffic data with location and identification; (5) personal control of subscriptions to public lists and unsolicited commercial communications. The essential criterion that allows data to be stored and processed by an organization is an effective agreement by the individual when providing his data. The policy covers all data sent over public networks in Europe and, therefore, also covers the data or services that originate outside Europe. In the light of these considerations, this paper provides a model for user control in pervasive environments formed by the adjustment of profile settings, where the pervasive environment context profile must be adjusted to control the privacy of users based on the characteristics of their lives and on rules laid down by Ref. [34]. 3.1 Criteria for a Model A model for the pervasive environment context was devised for this purpose and will be outlined below. In a ubiquitous environment, it is necessary to draw up criteria for user authentication. For this reason, we have defined the following controls: (1) Blocked: Access should be blocked to users for fixed or indefinite periods. (2) Guest: Access should be limited and controlled; Access can be made available for a fixed or indefinite period. Restrictions on services and sharing; Controlled privacy availability. (3) Basic: Controlled access. Sharing of resources and services is limited by the environment; The limiting factor will fall within a scale ranging from 1 to N, depending on the environment and its resources, his requires access to an ubiquitous data
A Privacy Taxonomy for the Management of Ubiquitous Environments
base where all available resources, in all pervasive environments, are registered; Sharing of location between other users of the pervasive environment. (4) Advanced: Access to all previous levels; Complete access to all resources and services in the environment. (5) Administrative: Access to all previous levels; Full control and management of the pervasive environment. These criteria will be assigned by the ubiquitous environment to the user that requests authentication in it. However, when dealing with external pervasive environments such as parks, squares or other public places, the user receives administrative access level, since no ubiquitous environment should exert dominance over a pervasive public environment. On the basis of these criteria, we conclude that it is necessary to adopt a middleware for the control and management of different environments and configurations. In Subsection 3.2, we describe the proposed architecture that will be used for this kind of middleware. 3.2 Middleware Architecture Model The (MW) overall architecture of the middleware supports all the necessary levels of control of the application, software and hardware, and will be based on the initial model proposed by Ref. [35]. In this model, there is a middleware focused on pervasive systems and divided into four layers: hardware, software, middleware and application. In this architecture, developed modules required for initial tests of a pervasive environment were validated using an OMAP platform [36]. Based on the above references, a few models were implemented in the tests that were carried out, namely a context management module, using a pervasive scheduling system as an application scenario, where users accessed their schedule in an intelligent and ubiquitous way. However, this architecture was not
1541
designed with applications involving WSNs or RFID. This observation can be confirmed at the time when protocols were implemented for pervasive environments with the aid of RFID and WSNs in the work conducted by Ref. [37]. In this application, it was possible to reduce energy consumption in wireless sensor networks with the widespread use of RFID tags in some parts of the pervasive environment and through a solution based on the ZigBee protocol. After identifying these deficiencies in the MW proposed in Ref. [35], we noticed the need for changes in the structure of middleware to give support for pervasive/ubiquitous applications and to manage RFID and WSN protocols and devices. Owing to these failings, the authors in Ref. [38] set out a new proposal for a MW platform capable of supporting the necessary technologies for pervasive and ubiquitous environments, with a focus on protocol management and reducing energy consumption, in accordance with the model shown in Fig. 2. The middleware proposed in Ref. [38] consists of four interconnected layers comprising the characteristics and requirements for the control of wireless sensor networks and RFID in pervasive/ubiquitous environments. One of its main features is the lowering of energy consumption caused by the reduction in the exchange of messages between the nodes and the base station. Thus, the middleware can be used to deal with hybrid problems involving pervasive/ubiquitous environments. The main idea is to allow individual devices to meet the needs of their users or the environment as a whole, by adapting to each environment and its underlying infrastructure to the best of their capability. The following is a description of each layer, as well as their characteristics. (1) Hardware Layer (HW): There are modules employed for handling the physical requirements necessary to deal with the physical devices and these are implemented in HW, as described:
1542
Fig. 2
A Privacy Taxonomy for the Management of Ubiquitous Environments
Hybrid middleware.
Network type: used to connect the middleware platform to the pervasive network. It also has specific features that will be used to assist and control the functionality of the device. I/O: used for communication and interfacing with users, environments and devices. Device drivers: pre-processing unit responsible for managing, storing and executing the minimum requirements for the operation of the devices. Acts as a trigger for the connection of physical devices registered in the system, such as MAC address specifications, IMEI (international mobile equipment identity), and Bluetooth, among others. (2) OS (operating systems) Layer: aims to handle the functions of operating systems for embedded systems. It is divided into two sub-layers: the device drivers sublayer manages the components of the physical layer and the embedded operating system sublayer manages the tasks of the application that runs on the device and determines the services provided by it, while also coping with the limitations of the system. Thus, every change that occurs at the operating system level in the user devices is dealt directly at this layer. (3) Protocols: responsible for carrying out the handling and management of the data protocols used in pervasive/ubiquitous environments along with the wireless sensor network and other devices such as RFID.
(4) MW (middleware) layer: is a set of components that assist in the integration and treatment of devices by the pervasive/ubiquitous network and carries out necessary services and makes other features that comprise the middleware architecture available. This layer consists of six primary modules: Communication: Integrates the device in the network and manages the communication of the device with other devices. Services: Carries out the management of services, environmental resources and devices for pervasive application. Another attribute is that it provides and controls the adaptation of new SW components. Adaptation: Responsible for the adaptation and management of users, services, devices, applications, communications and pervasive/ubiquitous environments. Security and privacy control: Responsible for handling the security environment, providing control and authentication services. The main purpose of this module is to manage and control the various types of sharing, by relating them to the privacy of the environment. Context: Helps to detect the context of the user and the environment. Monitoring: Provides environmental and device monitoring for the application, by reporting status, errors and problems.
A Privacy Taxonomy for the Management of Ubiquitous Environments
(5) AP (application) layer: is a module that has fragments of the applications running in the environment. This layer is responsible for performing all the necessary settings for any application to run in the environment. An example might be the provision of services made available by a coffeemaker in the pervasive environment where a particular user is located. (6) Intermediate layer: This layer is intended to connect and interact with all layers simultaneously, and its main objective is to establish a connection between two layers without necessarily connecting to others. For example, a sensorâ&#x20AC;&#x2122;s only purpose may be to alert the application to the occurrence of an event. In such a simple situation, the sensor may have its own self-managing operating system, and does not need to be connected with the operating system and other layers, and thus save resources on the platform. The treatment of privacy in the environment will be included in the middleware context module, since the context can be handled at the user, device, communication, services and environmental levels. Thus, it is not necessary to restructure the existing architecture, or allow the existing middleware
Fig. 3
Generic model for privacy.
1543
proposed in Ref. [38] to be used again. Basically, there will be a specification within the module that will be called â&#x20AC;&#x153;triggersâ&#x20AC;?. These triggers are inserted options that are used to activate the context module. They include the following types of options: environment, user, devices, communication, services, application and others. This enables the use of the middleware architecture in the generic privacy model that will be described in Subsection 3.3. 3.3 Generic Privacy Model The use of a mechanism for managing of privacy in environments pervasive and ubiquitous must meet the application requirements. In some scenarios, it is necessary to collect information from users to the system operation. These information should be treated legally and ethically because of the privacy of individuals. We propose a generic model for managing privacy in environments pervasive and ubiquitous as shown in Fig. 3. The proposed generic model contains several components for controlling the environments pervasive and ubiquitous, as described follows:
1544
A Privacy Taxonomy for the Management of Ubiquitous Environments
(1) Data base: This data base stores information about rules and definitions of user, devices and communications in environments pervasive and ubiquitous. This data base is like a single register containing all information necessary for the control and management of the privacy mechanism. This data base is linked to the data module, represented by double-headed arrows, where information is exchanged with the data module. (2) Controller module: This module receives the access requests and makes the control of the data base tables directly, according to the requirements and definitions of access and control of environment. It also performs requests for validations and updates the data base after the information has been returned to the module calculated and refined previously by the control module. These refinements are based on the characteristics and definitions in each management module. The data updated and set are returned to the requesting with his permission in accordance with the received variables. (3) Data module: This module performs all calculations of variables and parameters received from other modules, and generates a single output information for each processing run. (4) PRICRI: This module defines the rules and criteria of access, use, sharing, location, etc.. These rules can be added, changed, modified and/or replaced in accordance with the environment and with established rules. These definitions are handled individually by other modules that have individual characteristics and controls. The operation settings are preset for each environment and can have variations, such as the same user access the same room on the same day with different criteria defined according to the time of access. (5) PRIDEV: The module management and privacy control of device has as goal treat the data that is transferred by devices once that such devices may be of the environment itself and of other itinerant device. The management and control are related to the
characteristics of software and hardware of each device (size, weight, screen resolution, operating system, media, etc.). (6) PRICOM: This management control and communications privacy. This module defines the various forms of communication within the environments pervasive and ubiquitous, such as restrictions sign, type of adapter used to the controller accesses like in the environment of the real world, in which certain environments only have one type of communication. (7) PRIADA: This management control and adaptation module, which is responsible for processing information related to the adaptation of software and hardware in environment pervasive and ubiquitous. For example, treatment of content and media to be used in different devices by presenting differences in performance, functionality, communication or configuration. (8) PRISER: These modules service management environment. This module treats of the information about the availability of services to be used individually for each environment. For example, the information shared with other environments, such as the devices, communication types, location of users, environment features and components that interact with users. The definitions and rules for the use and availability of these services are inserted into the module environmental criteria in order to control access and management. (8) PRIHIS: This module stores and treats the information about the user history, environment, devices and other variables that may be added later with the goal of obtaining contextual in formation. The main feature is the use of information captured over a given period based on other sources of information, as for example, multiple tracks, context, etc.. (9) PRIPRO: This module is performed the transaction control on the management of user profile. Its main objective is control the information, previously defined by a search engine, which has only
A Privacy Taxonomy for the Management of Ubiquitous Environments
the purpose of distributing and direct the synthesized information to the next modules, in order to adapt appropriately to individual privacy based in the individual profile. (10) PRISEC: This module makes the control and management related to security, both the user as the environment. The module receives the parameters and related data encryption or other security-related settings treatments and forwards them to the requester according to the need for each situation. For example, when entering in a given environment the user may be blocked by situations that are beyond the criteria set within the same environment as the date and time allowed. (11) PRIENV: This module is registered the attributes related to the environment. Thus, with this information, it is possible check and manage what makes up the environment as well, their capacities and resources in order to share them to users who need according to your availability. In the next section, it shows an application scenario that takes account the model proposed.
4. Application Scenario In Ref. [32], the authors developed the Percontrol, a system that automatically manages and keeps track of user attendance. This system detects the entrance and exit of users within an academic or business environment. Percontrol also improves user discovery and localization service, within the local environment based on Bluetooth, Wi-Fi and RFID identifiers. However, the initial versions of Percontrol did not anticipate the use of WSNs or ANNs [39], such versions only intended to automatize student attendance tasks in classrooms. The application scenario shows the potential pervasive and ubiquitous computation for improving efficiency in workplaces. It also attempts to illustrate different possible perspectives that can have on a single pervasive scenario. This work proposes an extension of the work developed in Refs. [39, 40],
1545
increasing the pervasive functionalities available in this user tracking system, with the objective of increasing control over environmental conditions through user’s mobile devices. Using SunSpot wireless sensors [18] and Arduino kits [40], Percontrol can sense and manage the temperature and luminosity of an environment; and by using ANNs, the system can also attempt to adjust the values of these environmental properties to fit user preferences and the number of people in the environment, turning it into an intelligent location. The sequence diagram in Fig. 4 shows the primary inter actions between all parts of the system, as well as the messages exchanged since a user is detected until the environment adapts to its preferences. When the application detects the entrance of a device in the environment, a web service that manages the associations between users and devices is accessed. The device is identified through its BDA (bluetooth device address), Wi-Fi or RFID. The application maintains a module called BlueID which holds a list of all devices that were ever detected. Each time the application verifies the devices currently present in the environment, it performs a comparison with the previously stored list; newly detected devices generate an “entry’ event while missing devices are associated with an “exit” event. When accessed, the web service returns the username to the application, and also associated device resources and personal preferences through the HTTP protocol and an XML format message. The application also communicates with the SunSpot sensors to fetch the room temperature, luminosity, humidity or other environmental data that may be used at a later time. The following format was used to communicate with the sensor: messageID and sensorType. Both messageID and sensorType are numerical values. The messageID field is used to associate sensor response with the respective BlueID request, an important step since communication is asynchronous.
1546
Fig. 4
A Privacy Taxonomy for the Management of Ubiquitous Environments
Sequence diagram.
The sensorType represents the type of data being sent (luminosity, humidity, etc.); the “\n” character is used to mark the end of a message, while the “:” character is used as a data separator. The current environmental state is compared with user’s environmental preferences in order to decide the needed changes to be done. After a decision is reached, the environment sends commands to the actuator controllers, connected through USB to an operating computer, to change the environmental characteristics (e.g. turning the A/C unit on and change the room temperature). The extension of Percontrol’s functionalities translated into a more complex architecture as shown in Fig. 5. Initially, the prototype application and its respective transmitters were tested with a Windows operating system, an environment that benefited from the use of SunSpot sensors [39] and Arduino hardware [40]. There were many other advantages that led us to
choose the Arduino boards, namely the embedded input/output ports, low cost and strong modularity. The main idea behind the use of Arduinos was to test their viability for middleware development in pervasive environments, not excluding the possibility of having these boards completely replace several individual sensors for an integrated, single board solution connected to a computer. Fig. 6 illustrates the operation of Percontrol. From Fig. 6, it can be seen that a central controller is missing for allowing the exchange of profiles in the environment, according to user preferences. The main challenge here is controlling the number of parameters generated by the application while using WSNs and ANNs; the number of parameters increase with the amount of nodes and this means larger energetic and resource demands, as well as an more complex neural network processing, which may compromise the ANN’s response time.
A Privacy Taxonomy for the Management of Ubiquitous Environments
Fig. 5
Systemâ&#x20AC;&#x2122;s architecture.
Fig. 6
Functioning of percontrol.
An efficient monitoring of the network performance is necessary to guarantee a good quality of service. An example of this can be observed in the amount of time necessary to obtain information regarding a monitored environment; if it takes too long to obtain environmental information, this information may lose its value from an application perspective. Management of performance may provide means for the application to define proper quality metrics. These
1547
may be influenced by node density, exposition time, amount of dissipated energy and other factors. A mechanism that evaluates the level of importance of information is necessary for the management of quality of service. For example, a sensor detecting a temperature of 20 °C during spring is a normal occurrence, but the measurement of 50 °C under the same circumstances is an abnormal event, which would turn it into a relevant situation that would
1548
A Privacy Taxonomy for the Management of Ubiquitous Environments
require extra attention; it would implicate an artificial intelligence mechanism that could compare the abnormal value with other measured values by other sensors to see if the information is reliable and determine the proper course of action. Information that is of great importance to the normal function of the system should imply a greater effort for proper delivery. That is, energy consumed in communications should vary depending on the importance of the data. Another relevant management aspect concerns the installation of ad-hoc networks in unknown areas, where the behavior of wireless communications can be highly unpredictable, with high error-rates and considerable delays which may compromise the value of the information provided to the application. Performance management usually includes quality assurance, performance monitoring, control and analysis [41]. The QoS management process begins with the detection of performance degradation and ends when the source of the problem is ceased or removed. In between, the process has many intermediary stages of situation analysis [22]. Initially, there were used only 2 sunspot tests kits containing two wireless sensor nodes communicating with each base station connected to the computer via USB. Thus, it comes the need to conduct a comparative study of routing protocols for use in different environments composed of wireless sensor. To this end, several techniques exist to treat this problem and also allied service discovery, one of the most important, by the SLP (service location protocol) [18], which basically consists of maintaining a directory that contains the services available to whom it is offering them to. However, it is necessary to study thoroughly the operation of routing protocols in order to verify the protocol that best fits the pervasive control system, it is not in the scope of this work—the study of routing protocols. Therefore, for this work there were conducted only some tests to validate the survey and obtained results that demonstrate the feasibility of
work and their implementation and use with Percontrol, contributing to the improvement of the system and data so that other researchers can use it.
5. Example of Use and Preliminary Results One issue when having multiple users on the same system is the problem of concurrent data; e.g. the configuration of an air-conditioning unit may be influenced by every user that registers in this environment, since each user might have its own preferred temperature, and the temperature itself is general throughout the whole environment. In order to bypass this problem, the decision-making process for selecting the best “average” temperature must take into account the individual preferences from all users within the environment. A widely used solution [37, 42, 43], that has shown great results is the use of AI (artificial intelligence), in particular ANNs [44]. A neural network bases itself on real data that has occurred in the past and has been stored within the system for posterior access and use. The main objective of this work is not the choice of proper protocols or AI tools, but the creation of novel help mechanisms for Percontrol. Our choice for an AI mechanism dwelt on neural networks, while routing mechanisms were TCP/IP and ZigBee. These choices are supported by published works in routing protocols [18], artificial intelligence [43], and comparison and use of neural networks [27, 45]. The neural network loads the entire history of a device being handled within the environment, and uses its historical data as training, in order to identify decision patterns that were assumed in a recent past. Considering our air-conditioning example, these patterns include the temperature that each user wants for a certain environment and what temperature was actually used when all users were taken into consideration. This type of analysis is crucial for the network’s decision—making process. Fig. 7 presents part of the source code used to define the desired and assumed temperatures. These values are fixed for
A Privacy Taxonomy for the Management of Ubiquitous Environments
Fig. 7
Part of the neural network’s source code.
testing, but in a real scenario they are fetched from a data base or an archive. The code shown in Fig. 7 is used to train the neural network. After training phase the next step is to test the network to determine if it is well-suited to solve the problem of finding the ideal temperature using past event data; in order to perform the testing, a graphical interface was developed. The interface receives the values for current data and returns the ideal temperature estimation, as shown in Fig. 8. After the neural network’s training, we could
Fig. 8
1549
Training the ANN.
identify the network’s response time after a user enters the environment, as shown in Fig. 9. The Fig. 9 presents response times of the ANN for the cases with 1, 3 and 10 distinct users identified by the pervasive environment, where X axis represents the number of users and Y axis represents the elapsed time. For a single user, the ANN took 3 seconds to process the information contained in the user’s profile, returning an average temperature with a value equal to the one defined by the user (since it is just a single person). For three users, the neural network took 5
1550
Fig. 9
A Privacy Taxonomy for the Management of Ubiquitous Environments
Neural network’s performance and response time.
seconds to respond, and for ten users it took 13 seconds. In Fig. 10, a screenshot shows information on the users identified by the system, as well as on the devices associated with them. To perform the identification of different environments, we used an Arduino Duemilanove [40]. It is a microcontroller board based on ATmega328, with 14 digital input/output pins, 6 analog inputs, a 16 MHz crystal oscillator, a USB connection, a power jack, an ICSP header, and a reset button. The board contains the necessary assets to support the
Fig. 10
User identification screen.
microcontroller and its use is as simple as connecting it to a computer with a USB cable or powering it with an AC-to-DC adapter or battery. With this board, it was possible to detect devices via Bluetooth, Wi-Fi and, after being integrated with an appropriate card reader, RFID. The reader fetched a RFID card’s serial number that can be cross matched with the user’s registration on the data base. For this purpose, a RFID card reader model YHY502CTG was used in conjunction with the Arduino board. After obtaining the necessary application data and performing
A Privacy Taxonomy for the Management of Ubiquitous Environments
Fig. 11
Validation equipment.
the necessary adjustments, the system was validated using a didactic MultiPIC development Kit which possesses its own internal programmer. The didactic development Kit was connected to a stepper motor that simulated a ventilator. The stepper motor can be put in action with different speeds; initially, we defined 3 different speeds that corresponded to 3 different profiles. Fig. 11 shows a picture of the assembled device. The software used for simulating a ventilator with 3 speeds and controlling the board and stepper motor was developed in C language using the development software from Microchip MTLab, and transferred to the microcontroller with the IC-Prog 1.06C, the software used to compile the source code onto the MultiPIC Kit’s processor. On the performed tests, the ANNs computed the average temperature from the user profiles and used current environmental information from the SunSpot sensors to correctly manage the ventilation system. From these tests, we conclude that Percontrol managed the pervasive environment in a satisfactory manner and that the primary objectives of this research were met, although there is still much room for improvements.
6. Conclusion and Future Work There are several papers that cover one or two of the taxonomic topics related to pervasive and
1551
ubiquitous computing, but few describe how in the future they will give priority to the treatment of privacy in pervasive environments and not just to the elements that surround it. A primary goal of our study was to identify the related work to the management and control of the privacy in pervasive and ubiquitous context. According to the current literature, the main researches on control and management of privacy are about communication, applications and services, user and devices. This paper advances the state of art about pervasive environments and proposes a generic model to control and manage the privacy on these scenarios. The main focus of the model is the environment instead of only the users and their devices. A prototype was developed to test to validate the generic model of privacy. The results confirmed the viability of device detection with Wi-Fi, Bluetooth and RFID and an improvement over previous Percontrol versions. Nevertheless, there is still some latency in registering new devices on the system, which may be reduced by further adjustments of the parameters sent to the ANNs. This work represents a significant contribution since it covers different areas and technologies within pervasive computation. In the future work, several parameters and definitions will be implemented and tested, new models of privacy control for users, devices and environments will be considered.
Acknowledgments This work was developed with the support of the CNPq Brasil (Conselho de desenvolvimento Científico e Tecnológico), with resources provided by the SWE Program (Doutorado Sanduíche Programa Brasileiro Ciência sem Fronteiras) and by CAPES in the post-doctoral. Also, it is supported by the Project “Pesquisadores na Empresa Projeto GVWISEGVDASA-UNISINOS-CNPq Brasil”, the UFRGS (Federal University of Rio Grande do Sul), the SAPO/Portugal Telecom and the Department of Informatics Engineering of the University of Coimbra,
A Privacy Taxonomy for the Management of Ubiquitous Environments
1552
Portugal.
References [1] [2]
[3]
[4] [5] [6] [7]
[8]
[9]
[10]
[11] [12] [13]
[14]
[15]
M. Weiser, The Computer for the Twenty-First Century, Scientific American (1991) 94-10. M. Satyanarayanan, Pervasive computing: Vision and challenges, IEEE Personal Communications 8 (2001) 10-17. D. Saha, A. Mukherjee, Pervasive computing: A paradigm for the 21st century, Computer 36 (2003) 25-31. RFID Journal, http://www.rfidjournal.com (accessed April 2012). Wireless Sensor Network [Online], Scientific Research, http://www.scirp.org/journal/wsn (accessed May 2012). S. Warren, L. Brandeis, The right to privacy, Harward Law Review 4 (1890) 193-220. V.R.Q. Leithardt, D. Nunes, A.G.M Rossetto, C.O. Rolim, C.F.R. Geyer, J.S. Silva, Privacy Management Solution in Ubisuitous Environments Using Percontrol, Journal of Ubiquitous Systems and Pervasive Networks (JUSPN) 5 (2014) 21-28. J. Ye, L. Coyle, S. Dobson, P. Nixon, Ontology-based models in pervasive computing systems, Knowl. Eng. Rev. 22 (2007) 315-347. E. Moschetta, R.S. Antunes, M.P. Barcellos, More flexible degrees of collaboration, security and privacy in service discovery in ubiquitous environments, in: XXVI Brazilian Symposium on Computer Networks and Distributed Systems (SBRC 2008), Rio de Janeiro, Brazil, May 26-30, 2008. M. Ros, M. D’Souza, A. Postula, I. MacColl, Wireless outdoor personal area network using adaptive inquiry scanning for location-based services, Personal and Ubiquitous Computing 17 (2013) 387-397. J. Krumm, A survey of computational location privacy, Personal Ubiquitous Comput. 13 (2009) 391-399. A.R. Beresford, F. Stajano, Location privacy in pervasive computing, IEEE Pervasive Computing 2 (2003) 46-55. A. Gorlach, A. Heinemann, W.W. Terpstra, Survey on location privacy in pervasive computing, in: Privacy, Security and Trust within the Context of Pervasive Computing, The Kluwer International Series in Engineering and Computer Science, 2004, pp. 23-34. D. Lupiana, C. O’Driscoll, F. Mtenzi, Taxonomy for ubiquitous computing environments, in: First International Conference on Networked Digital Technologies, Ostrava, July 2009, pp. 469-475. A. Patwardhan, V. Korolev, L. Kagal, A. Joshi, Enforcing policies in pervasive environments, in: The First Annual International Conference on Mobile and Ubiquitous
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
Systems: Networking and Services, Aug. 2004, pp. 299-308. F. Zhu, M.W. Mutka, L.M. Ni, Service discovery in pervasive computing environments, IEEE Pervasive Computing 4 (2005) 81-90. K.E. Defrawy, G. Tsudik, Privacy-Preserving location-based on-demand routing in MANETs, IEEE Journal on Selected Areas in Communications 29 (2011) 1926-1934. R.Q.V Leithardt, C.F.R. Geyer, J.E.R. Tavares, Comparative study of routing protocols used in wireless sensor networks, in: IV Congresso da Academia Trinacional da Fronteira, Foz do Iguaçu, Brazil, 2009. R.C. Mickunas, J. Al-Muhtadi, P. Naldurg, G. Sampemane, M. Dennis, Towards Security and Privacy for Pervasive Computing [Online], http://srg.cs.uiuc.edu/gaia/papers/ towards-percomp-security.pdf (accessed February 2012). V.C.K. Sobral, J.B. Mangueira Mangueira, Especificando Privacidade em Ambientes de Computação Ubíqua. http://www.inf.ufsc.br/ bosco/ensino/ine6406 (accessed November 2011). C.O. Rolim, V.R.Q. Leithardt, A.G. Rossetto, T.F.M. dos Santos, A.M. Souza, C.F.R. Geyer, Six degrees of separation to improve routing in opportunistic networks, International Journal of UbiComp 4 (2013) 11-22. V. Pereira, J.S. Silva, J. Granjal, R. Silva, E. Monteiro, Q. Pan, A Taxonomy of Wireless Sensor Networks with QoS, in: 2011 4th IFIP International Conference on New Technologies, Mobility and Security (NTMS), Paris, Feb. 2011, pp. 1-4. N. Li, N. Zhang, S.K. Das, B. Thuraisingham, Privacy preservation in wireless sensor networks: A state-of-the-art survey, Ad Hoc Networks 7 (2009) 1501-1514. L. Kagal, T. Finin, A. Joshi, Trust-based security in pervasive computing environments, Computer 34 (2001) 154-157. K. Henricksen, R. Wishart, T. McFadden, J. Indulska, Extending context models for privacy in pervasive computing environments, in: Third IEEE International Conference on Pervasive Computing and Communications Workshops, Kauai Island, March 2005, pp. 20-24. V.R.Q. Leithardt, G.A. Borges, I.M. Carrera, A.G.M. Rossetto, C.O. Rolim, D. Nunes, S.J. Silva, C.F.R. Geyer, Mobile architecture for identifying users in ubiquitous environments focused on Percontrol, in: The Seventh International Conference on Mobile Ubiquitous Computing, Systems, Services and Technologies, UBICOMM 2013, Porto, 2013. pp. 145-151. C.O. Rolim, F. Schubert, A.G.M. Rossetto, V.R.Q.
A Privacy Taxonomy for the Management of Ubiquitous Environments
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
Leithardt, C.F.R. Geyer, C. Westphall, Comparison of a multi output adaptative Neuro-Fuzzy inference system (manfis) and Multi Layer Perceptron (mlp) in cloud computing, in: 29th Brazilian Symposium on Computer Networks and Distributed Systems, Paris, July 25-27, 2012. K. Shankar, L. Camp, K. Connelly, L. Huber, Aging, privacy, and home-based computing: Designing for privacy, IEEE Pervasive Computing 11 (2011) 46-54. J.T. Lehikoinen, J. Lehikoinen, P. Huuskonen, Understanding privacy regulation in ubicomp interactions, Personal Ubiquitous Comput. 12 (2008) 543-553. G. Iachello, J. Hong, End-user privacy in human-computer interaction, Foundation and Trends in Humn-Computer Interaction 1 (2007) 1-137. J. Bardram, R. Kjær, M. Pedersen, Context-Aware User Authentication—Supporting Proximity-Based Login in Pervasive Computing, in: UbiComp, Seattle, 2003, pp. 107-123. V.R.Q. Leithardt, C.O. Rolim, A.G.M. Rossetto, C.F.R. Geyer, M.A.R. Dantas, J.S. Silva, D. Nunes, Percontrol: A pervasive system for educational environments, in: 2012 International Conference on Computing, Networking and Communications (ICNC), Maui, Jan. 30-Feb. 2, 2012, pp. 131-136. A.K. Dey, Providing architectural support for building context-aware applications, Ph.D. Thesis, College of Computing, Georgia Institute of Technology, Atlanta, December 2000. Europe’s Information Society [Online]. http://ec.europa.eu/information/society/policy (accessed March 2012). R. Babbitt, J. Wong, C.K. Chang, Towards the modeling of personal privacy in ubiquitous computing environments, in: 31st Annual IEEE International Computer Software and Applications Conference, Beijing, July 24-27, 2007, pp. 695-699. Texas Instruments. Texas instruments [Online],
[37]
[38]
[39] [40] [41]
[42]
[43]
[44]
[45]
1553
http://focus.ti.com/pdfs/wtbu/ti_mid_whitepaper.pdf (accessed December 2011). R. Silva, V.R.Q. Leithardt, J.S.Silva, C.F.R. Geyer, J. Rodrigues, F. Boavida, A comparison of approaches to node and service discovery in 6lowPAN wireless sensor networks, in: Proceedings of the 5th ACM Symposium on QoS and Security For Wireless and Mobile Networks, New York, Oct. 2009, pp. 44-49. N. Li, N. Zhang, S. K. Das, B.M. Thuraisingham, Privacy preservation in wireless sensor networks: A state-of-the-art survey, Ad Hoc Networks 7 (2009) 1501-1514. Sun spot world, Oracle Labs, http://www.sunspotworld.com/ (accessed May 2013). Arduino brasil, http://www.arduino.com.br/ (accessed April 2013). A.A. Loureiro, J.M.S Nogueira, L.B. Ruiz, R.A. Mini de Freitas, E.F. Nakamura, C.M.S. Figueiredo, Wireless sensor networks, in: Brazilian Symposium on Computer Networks (SBRC), May, Porto Alegre, 2003, pp. 179-226. A.D.P. Braga, T.B. Ludemir, A.C.P.F. Carvalho, Artificial Neural Networks: Theory and Applications, LTC Publisher, Rio de Janeiro, Brazil, 2007. O. Ludwig Jr., C.M.E. Montgomery, Neural networks: Fundamentals and applications in C program, Rio de Janeiro: Editora Moderna LTDA Science, 2007. C.O. Rolim, A.G. M. Rossetto, V.R.Q. Leithardt, C.F.R. Geyer, Analysis of a Hybrid Neural Network as a basis for prediction mechanism Situation, in: Congress of the Brazilian Computer Society (CSBC 2012) Symposium on Pervasive Ubiquitous Computing, Brazil, 2012. R. Silva, J. Sá Silva, C. Geyer, V. Leithard, F. Boavida, Use of GPS and 6lowpan in mobile multi-sink wireless sensor networks—issues and feasibility, in: 8th International Information and telecommunication Technologies Symposium, IEEE R9 Latim America, 2009, pp. 154-160.
Journal of Communication and Computer 10 (2013) 1554-1565
Triangle Routing Problem in Mobile IP Sherif Kamel Hussein and Khaled Mohamed ALmustafa Department of Communication and Networks Engineering, Faculty of Engineering, Prince Sultan University, Riyadh 11586, King of Saudi Arabia
Received: October 14, 2013 / Accepted: November 16, 2013 / Published: December 31, 2013. Abstract: Mobile Internet Protocol is a recommended Internet protocol designed to support the mobility of a user (host). Host mobility is becoming important because of the recent blossoming of laptop computers and the high desire to have continuous network connectivity anywhere the host happens to be. The development of Mobile IP makes this possible. The traditional Mobile IP specification forces all packets forwarded to the MN (mobile node), to be routed via HA (home agent), which often leads to Triangular routing, which in turn causes data transmission delay and wastes network resources. This paper discusses means of resolving the triangle routing problem, it introduces some of the recent route optimization schemes that have been used to solve that problem. Key words: Mobile IP, triangle routing problem, route optimization.
1. Introduction ď&#x20AC; Mobile IP is an open standard, defined by the IETF (Internet Engineering Task Force) RFC 2002, that allows users keep the same IP address, stay, connected, and maintain ongoing applications while roaming between networks, given that any media that can support IP can support Mobile IP. Efforts were made to enhance the standard protocol and to be able to achieve data transmission within the wireless infrastructure. However, in trying to achieve this goal, many problems have emerged and solutions are evolving [1]. The key feature of Mobile IP design is that all required functionalities for processing and managing mobility information are embedded in well-defined entities, the HA (home agent), FA (foreign agent), and MNs (mobile nodes). The MN is a host or router that can change its location from one link to another without changing its IP address or interrupting existing services. The home agent is a router with an Corresponding author: Sherif Kamel Hussein, assistant professor, research fields: mobile communication, automation and wireless control. E-mail: skhussein@psu.edu.sa.
interface on a mobile nodeâ&#x20AC;&#x2122;s home link that intercepts packets destined for the home address; it tunnels packets to the mobile nodes most recently reported care-of-address. The foreign agent is a router on a mobile nodeâ&#x20AC;&#x2122;s visited network that provides routing services to the mobile node while it is registered [2-4]. Triangle routing problem is considered as one of the main problems facing the implementation of Mobile IP such as, when a CN (communicating node) sends traffic to the mobile node, packets first get to the home agent, which encapsulates these packets and tunnels them to the foreign agent. The foreign agent de-tunnels the packets and delivers them to the mobile node. The route taken by these packets is triangular in nature, and the most extreme case of routing can be observed when the communicating node and the mobile node are in the same subnet [5, 6]. In recent literature, many protocols have been invented to solve the triangle routing problem. In this paper, we introduce some of recent route optimization schemes that are used in solving the conventional triangle routing problem in Mobile IP [7-14]. The most common schemes that will be discussed in this
Triangle Routing Problem in Mobile IP
paper are: Forward tunneling and binding cache, dynamic address allocation, bidirectional route optimization and finally the ISP PoPs (Internet Service Provider Points of Presence) [7-10]. The paper is divided into five sections. The second section, examines the modifications to IP protocol to accommodate wireless access to the internet (Mobile IP). The third section, introduces the concept of the triangle routing problem in Mobile IP. The fourth section, details a survey for some previous recent protocols proposed for optimizing the triangle routing problem plus their drawbacks. The fifth section provides conclusion and future work.
2. Mobile IP The main problem in the process of introducing mobility to the Internet is IP addressing. The IP address is a unique address for each network access point (e.g. in a router, a terminal, and so forth). Furthermore, the IP address is used for routing packets in the intermediate routers between the source and the destination, so the problem for mobility in the Internet is how to handle the mobile terminal’s IP address and routing information when the mobile host makes handoff between two wireless access points (e.g. base stations) or when it roams between two network domains (i.e. between two network operators). A solution to this problem is provided through the Mobile IP Protocol [15]. This protocol provides mobility, also, it is transparent to the transport and higher protocol layers. Its implementation does not require any changes in the existing nodes and hosts on the Internet. 2.1 Mobile IP Definition Mobile IP is a modification to IP that allows nodes to continue to receive datagrams no matter where they happen to be attached to the Internet. It involves some additional control messages that allow the IP nodes involved to manage their IP routing tables reliably. Scalability has been a dominant design factor during
1555
the development of Mobile IP, because in the future a high percentage of the nodes attached to the Internet will be capable of mobility [5, 16]. 2.2 Mobile IP Terminology The Mobile IP terminology illustrates the following: (1) Mobile node: A host or router that changes its point of attachment from one network or subnetwork to another; (2) Ha (home address): an IP address that is assigned for an extended period of time to a mobile node in the Home Network; (3) HA (home agent): A router on a mobile node’s Home Network which tunnels datagrams for delivery to the mobile node when it is away from home, and maintains current location information for the mobile node; (4) Home network: A network, possibly virtual, having a network prefix matching that of a mobile node’s home network; (5) FA (foreign agent): A router on a mobile node’s visited network which provides routing services to the mobile node while registered. The FA de-tunnels and delivers datagrams to the mobile node; (6) Foreign Network: Any network other than the mobile node’s home network; (7) CoA (care-of-address): The termination point of a tunnel toward a mobile node, for datagrams forwarded to the mobile node while it is away from home; (8) CN (correspondent node): A peer with which a mobile node is communicating, it may be either mobile or stationary; (9)Link: A facility or medium over which nodes can communicate at the link layer. A link underlies the network layer; (10) Node: A host or a router; (11) Tunnel: The path followed by a datagram while it is encapsulated; (12) Virtual network: A network with no physical
Triangle Routing Problem in Mobile IP
instantiation beyond its router (with a physical network interface on another network); (13) Visited network: a network other than a mobile node’s home network to which the mobile node is currently connected; (14) Visitor list: The list of mobile nodes visiting a foreign agent; (15) Mobile binding: The association of home network with a care-of-address, along with the remaining lifetime of that association. 2.3 Operation of Mobile IP Mobile IP is, in essence a way of doing the following three relatively separate functions, agent discovery, registration and tunneling [15, 16]. 2.3.1 Agent Discovery The discovery process in Mobile IP is very similar to the router advertisement process defined in ICMP. For the purpose of discovery a router or other, network node that can act as an agent periodically issues a router advertisement ICMP message with an advertisement extension. Fig. 1 shows a general abstract format for the Mobile IP agent advertisement message. The router advertisement portion of the message includes the IP address of the router. The advertisement extension includes additional information about the router’s role as an agent. A mobile node listens for these agent advertisement messages. The mobile node must compare the network portion of the router’s IP address with the network portion of its own home network. If these network portions do not match, then the mobile node is on a foreign network [15, 16]. MH 1. Requests service
FA
Mobile IP registration overview.
Fig. 1
Agent advertisement extension
Optional prefixlength extension
General Mobile IP agent advertisement message.
FA = Foreign Agent MH = Mobile Host HA = Home Agent 2. FA relays
FA
3. HA accept
FA Fig. 2
ICMP router Advertisement
FA
4. FA relays Status to MH
The prefix extension that follows the agent advertisement extension is used to indicate the number of bits of network prefix that apply to each router address listed in the ICMP router advertisement portion of the agent advertisement 2.3.2 Registration Once a mobile node has recognized that it has transferred on a Foreign Network and has acquired a care-of-address, it needs to alert a home agent on its Home Network and requests that the home agent forwards its IP traffics. The registration process involves four steps as shown in Fig. 2: (1) The mobile node requests the forwarding service by sending a registration request to the foreign agent that the mobile node wants to use; (2) The foreign agent relays this request to the mobile node’s home agent; (3) The home agent either accepts or denies the request and sends a registration reply to the foreign agent; (4) The foreign agent relays this reply to the mobile node. If the mobile node uses a co-located care-of-address; it registers directly with its home agent, rather than going through a foreign agent. The registration process involves two steps: (1) The mobile node sends a registration request to the home agent. (2) The home agent sends a registration reply to the mobile node that grants or denies the request.
HA HA
Time
1556
Triangle Routing Problem in Mobile IP
Mobile IP registration messages use UDP (user datagram protocol). The overall data structure of the registration messages is shown in Fig. 3 [15, 16]. 2.3.3 Tunneling Once a mobile node is registered with a home agent, the home agent must be able, to intercept IP datagrams sent to the mobile node’s Home Network so that these datagrams can be forwarded via tunneling. In the most general tunneling case, illustrated in Fig. 4, the source, the encapsulator, the decapsulator and the destination are separate nodes. The encapsulator node is considered the entry point of the tunnel, while the decapsulator node is considered the exit point of tunnel. Multiple source-destination pairs can use the same tunnel between the encapsulator and decapsulator [15, 16]. Three options for encapsulation (tunneling) are available for use by the home agent on behalf of the mobile node. (1) IP-ln-IP Encapsulation: To encapsulate an IP datagram, an outer IP header is inserted before the datagram’s existing IP header, as shown in Fig. 5a. The outer IP header source address and destination address identify the endpoints of the tunnel. The inner IP header source address and destination address identify the original sender and recipient of the datagram respectively. The inner IP header is not changed by the encapsulator, and remains unchanged during its delivery to the tunnel exit point. (2) Minimal Encapsulation: To encapsulate an IP datagram using minimal encapsulation, the minimal forwarding header is inserted into the datagram, as shown in Fig. 5b. The IP header of the original datagram is modified and the minimal forwarding header is inserted into the datagram after the IP header, followed by the unmodified IP payload of the original datagram (for example, transport header and transport data). No additional IP header is added to the datagram. (3) GRE (general routing encapsulation): GRE is
1557
more general than the other protocols described earlier. It can encapsulate numerous other protocols besides IP. The entire encapsulated packet has the form presented in Fig. 5c. 2.4 Mobile IP Operation Sequence With the three relatively separated functions, agent discovery, registration and tunneling, a rough outlines of the operation of Mobile IP Protocol is described as shown in Fig. 6 [5]. IP Header fields Fig. 3
UDP header
Mobile IP message header
Extensions…
General Mobile IP registration message format. Tunneling
Encapsulation
Decapsulation
Destination
Source Fig. 4
General tunneling.
Tunnel endpoints
Original IP header
Original IP header
Inner IP header
Original IP payload
Original IP payload
Other headers (optional) (a)
Tunnel endpoints
Outer IP header
Delivery Header
Original IP header Original IP payload Destination IP Address Minimal Encapsulation header (b)
Original IP payload
GRE Header Packet Payload (c) Fig. 5 Types of encapsulation: (a) IP-In-IP Encapsulation; (b) Minimal Encapsulation；General Routing Encapsulation.
Triangle Routing Problem in Mobile IP
1558
agent through the exchange of a registration request and registration reply message, possibly by way of a foreign agent. (6) Datagrams sent by the correspondent node to the mobile nodeâ&#x20AC;&#x2122;s Home Network are intercepted by its home agent, tunneled by the home agent to the mobile nodeâ&#x20AC;&#x2122;s Care-of- Address, received at the tunnel endpoint (either at a foreign agent or at the mobile node itself), and finally delivered to the mobile node. (7) In the reverse direction, datagrams sent by the mobile node may be delivered to their destination using standard IP routing mechanisms, without necessarily passing through the home agent.
(1) Mobility agents (FAs (foreign agents) and home agents) advertise their presence via agent-advertisement messages. A mobile node may optionally solicit an agent advertisement message from any local mobility agents by using an agent solicitation message. (2) A mobile node receives an agent advertisement and determines whether it is on its Home Network or a Foreign Network. (3) When the mobile node detects that it is located on its Home Network, it operates without mobility services. If returning to its Home Network from being registered elsewhere, the mobile node deregisters with its home agent through a variation of the normal registration process. (4) When a mobile node detects that it has moved to a Foreign Network, it obtains a CoA (care-of-address) on the Foreign Network. The care-of-address can either be a foreign agent care-of-address or a co-located Care-of- Address. (5) The mobile node, operating away from home, then registers its new care-of-address with its home
2.5 Security in Mobile IP Security is an increasing concern in the design of mobile networking protocols and systems [17-19]. Authentication is critical to authorizing operations indicating the mobile nodeâ&#x20AC;&#x2122;s new point of attachment. The three common security measures in today's Internet that affect the mobile networking are:
1. Send agent advertisement Foreign Network
1. Send agent advertisement
FA 4.
5.b. Reg. Request HA Mobility binding
MN
5.c. Reg. Reply 3. 2. Received agent advertisement
7. Datagrams sent from MN to CN Fig. 6
Mobile IP operation sequence.
CN
6.
Data to Mobile Node Intercepted by HA
5.a. Reg. Request
Triangle Routing Problem in Mobile IP
(1) Firewalls: The existence of firewalls is an unfortunate reality in today’s Internet. Firewalls perform the function of discriminating against IP datagrams transiting the enterprise’s border routers to protect the computing assets of the enterprise against attack by the millions of Internet computers not associated with the enterprise [15, 16]. (2) Border Routers: Border routers can be configured to discard incoming datagrams that seem to emanate from internal computers. The philosophy is to prevent computers in the external Internet from spoofing (using the address of) internal computers [15, 16]. (3) Ingress Filtering; It has recently been proposed that border routers at the periphery of an administrative domain (for instance, supporting an ISP (internet service provider)) carefully discard datagrams that seem to emanate from an address external to the administrative domain. This feature is called ingress filtering [15, 16].
3. Triangle Routing Problem One of the basic problems facing the implementation of Mobile IP is the triangle routing problem, since all the traffics between correspondent node and mobile node should have to pass through a longer path than the normal one. This section introduces the definition and the drawbacks for the triangle routing problem, also, it details some recent papers for route optimization scheme. 3.1 Triangle Routing Definition Triangle routing problem is considered as one of the problems facing the implementation of Mobile IP. When a correspondent node sends traffics to mobile node, the following sequence must be done: Packets first get the home agent; Home agent encapsulates these packets and tunnels them to the foreign agent; The foreign agent de-tunnels the packets and delivers them to the mobile node.
1559
As shown in Fig. 7, the route taken by these packets is triangle in nature, and the most extreme case of routing can be observed when the correspondent node and mobile node are in the same subnet [7, 16]. 3.2 Triangle Routing Drawbacks Conventional Mobile IP scheme allows transparent interoperation between mobile nodes and their correspondent nodes, but forces all datagrams for a mobile node to be routed through its home agent. Thus, datagrams to the mobile node are often routed along paths that are significantly longer that optimal. This indirect routing can significantly delay the delivery of the datagrams to mobile nodes, and it places an unnecessary burden on the networks and routers along its path through the internet. So we can summarize the Triangle Routing drawbacks as follow: Increases the delays per packet in datagrams transferred to the mobile node; Waste of network resources; Home agent bottle neck; Delimits the scalability of Mobile IP protocol.
4. Triangle Problem’s Previous Solutions There have been attempts to eliminate or address the triangle routing problem in Mobile IP [7-14]. This section introduces some of recent attempts dedicated for solving the triangle routing problem [7-10]. Datagram MN-CN CN
MN Datagram MN-CN 3 1
Datagram MN-CN
FA 2 HA
Tunneled datagram
Fig. 7 Illustration of the triangle routing problem in Mobile IPv4.
Triangle Routing Problem in Mobile IP
handoff when the mobile node moves and registers with a new foreign agent [20, 21]. It provides a means for the previous foreign agent to be notified of the mobile nodeâ&#x20AC;&#x2122;s new mobility binding allowing data grams in flight to the mobile nodeâ&#x20AC;&#x2122;s foreign agent to be forwarded to its new care-of-address as shown in Fig. 9.
4.1 Route Optimization by Forward Tunneling and Binding Cache Route Optimization Protocol in Fig. 8 was developed to solve the Triangular Routing Problem, by allowing each host to maintain a binding cache for a mobile host wherever it is. When sending a packet to a mobile node, the following sequence must be taken: (1) If the sender has a binding cache containing the care-of-address of the mobile node, it will deliver the packets directly towards the mobile node. (2) If the sender has no binding information the first packets should be destined at first to the HA (home agent). (3) Home agent encapsulates the packets and sends them to the foreign agent. (4) Foreign agent decapsulates the packets and send them to the mobile node. (5) Binding information is transferred from the home agent to the source node for the further correspondences in the future, such that the next packets should be routed directly to the Foreign Network. If mobile node sends packets to the source node, the packets will be transferred directly from the mobile node to the source node. This Route Optimization scheme provides a smooth
Home Agent
As a result of simulated comparison between the original Mobile IP scheme and the improved scheme which use forward tunneling and binding cache. It has been proved that, the transmission time (delay) between the correspondent node and the mobile node is reduced because of the shortest path to reach the mobile node. The traffic and control signals over the network have been decreased [7]. 4.2 Route Optimization Using Dynamic Address Allocation in Mobile IP This technique proposes an extension to the Mobile IP architecture. In this scheme one MS (mobile station) is to handle two IP addresses between internet and intra-domain, one is called CA (current address) and another one is called RA (register address). LA (location agent) is a router responsible for translating both addresses between internet and intra-domain. RA is used for packets routing in internet; CA is used for packets in intra-domain. MA (mobile agent) is router
Packets from source node Destined to the MN with no binding cache for the sender 2 5 Binding update Next packets from source node routed directly to the foreign network
3 Encapsulation Foreign Agent Fig. 8
Route optimization.
Source Node
1 If sender has a binding information
1560
4 Decapsulation
6 Packets back to the source directly
Mobile Node
Triangle Routing Problem in Mobile IP
1561
Register with Home Agent Internet work Notify previous Foreign Agent Foreign Agent
Previous Foreign Agent
Register with foreign agent Mobile host Fig. 9
Smooth handoff during registration.
Hard handoff scheme is proposed to be used with this technique. Also a “packet retransmission” scheme is used to avoid packet loss while hard handoff, in which every MA should have a buffer to store the downlink packets transmitted to MS. After MS handoff, old MA would retransmit packets which are stored in its buffer to new MA which delivers them to MS. By evaluating the performance comparison between the Mobile IP scheme and the dynamic address allocation scheme, it has been found that the transmission time taken between the CN and MS takes a longer Downlink path in case of Mobile IP scheme than the dynamic allocation scheme in which the transmission time equal to the time taken between the CN and MA plus the time taken between MA and MS. Also the traffic would increase obviously in the Mobile IP scheme. Comparatively, dynamic allocation scheme would not increase any extra traffic [8].
on a MS’s current network which delivers packets to MS, it has a functionality similar to foreign agent and home agent. Considering a packets routing scheme between MS and CN, the packets downlink and uplink of the proposed architecture are described in Fig. 10 as following: When a MS sends a packet to CN, the packet routed first to LA by using CA. When LA receives packet, it will use the CA of MS to check relative address of MS. LA uses the RA instead of CA and retransmits the packet to CN. This sequence is called packet Up Link sequence. In case of packet down link, when a CN sends a packet to MS, the packet is routed to LA by using RA first. When LA receives the packet, it will use the RA of MS to check the relative CA of MS. LA uses the CA instead of RA and retransmits the packet to MS. Internet
Up Link Direction 1.c 2.a
Correspondence Node
1.b
Location Agent Mobile Agent
Fig. 10
Dynamic allocation mobile IP architecture.
Down Link Direction 2.b 1.a MS Mobile Agent
2.c
Triangle Routing Problem in Mobile IP
1562
4.3 Bi-directional Route Optimization in Mobile IP over Wireless LAN This technique is proposed to support symmetric bidirectional route optimization in Mobile IP considering ingress-filtering routers [22]. Subnet-based direct tunneling techniques are proposed to improve the routing efficiency for Mobile IP and a binding optimization technique to reduce the handoff latency for mobile nodes. An enhanced correspondent agent was introduced to collaborate with the home agent and the foreign agent to support these techniques. Fig. 11 presents the overall design of the bidirectional route optimization. It is used to address the issues of Triangle Routing and ingress filtering in Mobile IP. The design introduces a correspondent agent which maintains the binding cache and intercepts all packets sent to and from the correspondent nodes. Symmetrically, a foreign agent, at the other end of the optimized route or tunnel, maintains a tunneling cache for bidirectional route optimization. An entry of the tunneling cache indicates that a correspondent node or Correspondent Network supports Route Optimization and direct tunneling, so a foreign agent can directly tunnel a packet received from a mobile node to the correspondent node that matches a tunneling cache entry. So a home agent or a foreign
agent must be able to distinguish the traditional correspondent node from the enhanced correspondent agent, and the enhanced correspondent agent must also be able to distinguish a mobile node from a usual Stationary Node. From the preliminary results of the simulation over a WLAN it has been found that; using Bi-directional route optimization not only improves the routing efficiency but also reduces the handoff latency for mobile node. Also, the packet transmission time and the traffic can also be reduced through direct tunneling. Further simulation is under development [9]. 4.4 Route Optimization Using Internet Service Provider Points of Presence ISP-POPs The basic idea in optimizing Triangle Routing is to get the home agent as close as possible to the mobile node, when the mobile node no longer in its Home Network. This is achieved by shifting the home agent into the ISPâ&#x20AC;&#x2122;s Domain. The ISPâ&#x20AC;&#x2122;s network can be made Mobile IP "aware" by enhancing ISP-PoPs (Points of presence) and by creating a virtual network composed of PoPs to distribute the state information about the mobile node at the original home agent to all PoPs. This ensures that no matter where the mobile node is, the home agent is just a PoP away [10]. Fig. 12
Home Network Home Agent Correspondent Agent
Direct Tunneling Subnet-based Binding Opt.
Binding cache Correspondent Node Correspondent Network Fig. 11
Architecture of bi-direction route optimization.
Foreign Agent
Tunneling cache Mobile Node Foreign Network
Triangle Routing Problem in Mobile IP
1563
CN POP1
Seattle, WA USA
Moscow Russia
POP2
ISP POP3 MN at home FA Tokyo, Japan MN away Fig. 12
Simplistic view of a global network.
offers a simplistic view of a global network showing the position of CN, mobile node, PoP1, PoP2, PoP3 and ISP. The following paragraphs describe events that occur in order to set up the stage for successful CN to mobile node communication. (1) Registration with the home agent; The mobile node registers with the closest home agent, in this case, PoP1; PoP1 informs PoP2 and PoP3 about mobile nodes intent to be mobile through the PVN (PoPs virtual network) as shown in Fig. 13a (based on Fig. 12). (2) Registration with the foreign agent; The mobile node registers with the foreign agent, seeking mobile services in the new network; The foreign agent gets in touch with the nearest home agent, in this case PoP3 in order to authenticate the mobile node; PoP3 now knows that to reach the mobile node, it only needs to reach the foreign agent. So it creates an explicit routing entry mapping the mobile node
address with the foreign agent address. PoP3 broadcasts this information to all other PoPs, so that they may also do the same as shown in Fig. 13b (based on Fig. 12) ; On receiving a successful reply for the mobile node authentication message, the foreign agent creates an association between the original mobile node address and its current point of attachment in the subnet. It uses this association to replace the destination IP address in incoming packets with the current point of attachment address. Similarly, the source address in outgoing packets from the mobile node is replaced with the original mobile node address. (3) CN needs to communicate with mobile node; Packets addressed to the mobile node’s address can be routed by PoP3, since it has an explicit routing destination (FA) for such packets; Normal Internet routing gets these packets to the FA, where a destination address swapping occurs, in which the original MN address is swapped with its current address; Reg (1)
MN
Reg Reply
Reg Reply
PoP2
PVN update (2)
Auth Reply (4) PoP3
PoP1(HA)
Fig. 13
FA
MN
Reg (1)
PoP2
MN Auth (2) PoP3
PVN update (3)
(a) (b) HA and FA registration with PVN update messages: (a) HA registration; (b) FA registration.
PoP1
1564
Triangle Routing Problem in Mobile IP
The IP protocol stack at the mobile node can now receive the packets originally destined for the mobile node address. (4) mobile node gets back to its Home Network; The mobile node de-registers with the closest home agent, in this case PoP1; PoPl informs its peers, and all state information regarding the mobile node is purged; FAs purge their associations based on the lifetime of the association; As the result of simulation comparison applied between the two approaches, the Conventional Mobile IP framework and the new proposed Mobile IP framework. It has been found that applying the new framework for Mobile IP is best suited to large ISPs with a large topological reach. It increases the TCP throughput to mobile node by almost double that in that traditional Triangle Routing case. Also the transmission time (delay) has been reduced through the new framework. 4.5 Previous Route Optimization Schemes Drawbacks in Mobile IP The great effect of using the route optimization schemes is to minimize the transmission time (delay) between CN (correspondent node) and MN (mobile node) because of the shortest path to reach mobile node and also to reduce the traffic and control signals over the network. The drawbacks of the most Route Optimization techniques are classified as follows: (1) Rigid requirements for an authentication of the clamed care-of-address especially when both of mobile node and correspondent node are in a different IP networks; (2) Expensive hardware devices needed for the Route Optimization functions; (3) Increase the amount of traffics over the network; (4) Increase the rate of buffering and storage buffers; (5) Increase the rate of blocking especially when
the number of connections to mobile nodes is increased which results in increasing in the transmission time between correspondent node and mobile node.
5. Conclusions and Future Work In this paper, we introduced the definition, the operation and the security in Mobile IP Protocol. The discussion includes the three main functions for the Mobile IP operation; agent discovery, advertisement and the tunneling procedures. A description for the triangle routing problem in Mobile IP plus some recent researches with their drawbacks has been discussed. Forward tunneling, dynamic agent allocation, bidirectional route optimization and the ISP-PoPs were proposed as recent researches that provide a solution for the triangle routing problem in conventional Mobile IP Protocol. As future work, we shall introduce a new route optimization scheme for solving the triangle routing problem in conventional Mobile IP protocol. This scheme will be proposed to recover most of the drawbacks for the previous route optimization schemes; by increasing the level of authentication, decreasing the hardware devices needed, decreasing the amount of traffic over the network and finally decreasing the rate of buffering and blocking especially when the number of connections to mobile nodes are increased.
References [1] [2] [3]
[4]
[5]
C. Smith, D. Collins, 3G Wireless Networks, McGraw-Hill, United States, 2002. A. Jamalipour, The Wireless Mobile Internet, John Wiley & Sons Ltd., England, 2003. J.P. Nesser II, Survey of IPv4 addresses in currently deployed IETF standards, IETF Internet-Drafts, August 2001. S.G. Choi, R. Mukhtar, J.K. Choi, M. Zukerman, Efficient Marcro mobility management for GPRS IP networks, Journal of Communications and Networks 5 (2003) 55-64. W. Stallings, Wireless Communications and Networks, Prentice Hall, New Jersey, United States, 2002.
Triangle Routing Problem in Mobile IP [6] [7] [8] [9]
[10]
[11]
[12]
[13]
T. Janevski, Traffic Analysis and Design of Wireless IP Networks, Artech House Inc., Boston, London, 2003. C. Perkins, IP Mobility Support for IPv4, IETE RFC 3344, August 2002. W. Wu, W.S. Chen, F.R. Young, H.E. Liao, Dynamic Address Allocation in Mobile IP, Nov. 1999. C.H. Wu, A.T. Cheng, S.T. Lee, J.M. Ho, D.T. Lee, Bi-directional route optimization in mobile IP over wireless LAN, in: 2002 IEEE 56th Vehicular Technology Conference, Vancouver, Canada, 2002. An Efficient, Global Mobile IPv4 Routing Frame Work using Internet Service provider, Point of Presence ISP-PoP [Online], http:\\networks.ecse.rpi.edu/papers/mip.pdf. C. Kumar, N. Tyagi, R. Tripathi, Performance of mobile IP with new route optimization technique, in: IEEE International Conference on Personal Wireless Communications, New Delhi, Jan. 23-25, 2005, pp. 522-526. D. Badami, N. Thanthry, T. Best, R. Bhagavathula, R. Pendse, Port address translation based route optimization for mobile IP, in: IEEE 60th Vehicular Technology Conference, Los Angeles, Sept. 26-29 2004, pp. 3110-3114. R. Vadali, J.H. Li, Y.Q. Wu, G.H. Cao, Agent based
[14]
[15] [16] [17] [18] [19]
[20] [21]
[22]
1565
route optimization for mobile IP, in: IEEE VTS 54th Vehicular Technology Conference, Atlantic City, Oct. 7-11 2001 pp. 2731-2735. Q. Gao, A. Acampora, A virtual home agent based route optimization for mobile IP, in: IEEE Wireless Communication and Networking Conference, Chicago, Sept. 23-28 2000, pp. 592-596. C.E. Perkins, Mobile IP, IEEE Communications Magazine 35 (1998) 84-99. C.E. Perkins, Mobile IP: Design Principles and Practices, Addison-Wesley, United States, 1998. S. Vaarala, Mobile IPv4 traversal across, IPSec Based VPN Gateways, IETF Internet-Drafts, June 2003. H. Hansen, IPSec and Mobile IP in Mobile Ad-Hoc Networking, April 2000. J. Arkko, V. Devarapali, F. Dupont, Using IPSec to protect mobile IPV6 signaling between mobile node and home agents, IETF Internet-Drafts, February. 2003. B. Ayani, Smooth handoff in Mobile IP, Master Thesis, University of California in Berkeley, May 2005. T. Hiller, J. Kempt, P.J. Mccann, A. Singh, H. Soliman, S.Thalanany, Low latency handoffs in mobile IPv4, IETF Internet-Drafts, December 2003. G. Montenegro, Reverse Tunneleng for Mobile IP, RFC, USA, January 2001.
Journal of Communication and Computer 10 (2013) 1566-1572
Geographical Monitoring of Electrical Energy Quality Determination: The Problems of the Sensors Maurizio Caciotta, Fabio Leccese, Sabino Giarnetti and Stefano Di Pasquale Science Department, University of Roma Tre, Roma 00146, Italy
Received: November 2, 2013 / Accepted: December 15, 2013 / Published: December 31, 2013. Abstract: The problem of the electrical energy price, evaluated considering its quality, has been analyzed in this article. The electrical energy quality is directly defined by the customers as an output of a complex procedure which uses as inputs both the power q uality technical parameters and the human experience. The procedure shows as the accuracy in the quality definition is strongly related with the accuracy of the sensors used for the power quality measurements, especially regarding their synchronization. An analysis of the synchronization of the sensors placed on a spread territory is also made. A solution for a centralized control of the sensors time course is also suggested. Key words: Power quality, perceived power quality, power probes, timing of distributed probes.
1. Introductionď&#x20AC; To evaluate the voltage of the mains, The Electric and Electronic Measurement Laboratory of the Science Department of Roma Tre University installed some measurements probes in two Italian cities, Rome and Palermo. The places of the cities on the Italian land are depicted in Fig. 1. The placements of the probes in the urban areas are depicted in Fig. 2, both for Rome and for Palermo. They are collocated inside telephone exchanges of the most important Italian telecommunication provider: Telecom Italia. Each probe picks up the mains voltage at frequency of 25 kHz. The aim of these devices is to evaluate the time course of the technical parameters defined in the CEI EN 50160 (also called PPQ (power quality parameters)), which is the norm that defines all the characteristics of the electrical energy supplied by public distribution systems [1]. The norm defines thirteen technical parameters
which are locally determined by the single probes and are then interpolated on the whole city territory. The aim of the article is to describe a complex procedure which, using as inputs both the power quality technical parameters and the human experience, allows to define the electrical energy quality directly by the customers. The procedure shows as the accuracy of the quality definition is strictly joined to the accuracy of the sensors used, especially regarding their synchronization placed on a spread territory. The paper is organized as follows: after a brief introduction, the Section 2 explains how is made the probe and how it works; Section 3 explains the timing problem; in Section 4 we discuss of the economic importance of the application; Sections 5 describes the frequency measurements algorithm and the related accuracy; Section 6 shows some tests and measurements obtained from the probes in field; Section 7 describes the future improvements while Section 8 is the conclusion.
2. The Probe Block Scheme Corresponding author: Fabio Leccese, Ph.D., assistant professor, research fields: power quality, signal analysis, sensor networks and smart grids. E-mail: leccese@uniroma3.it.
The Probe block scheme is shown in Fig. 3.
Geographical monitoring of Electrical Energy Quality determination: The Problems of the Sensors
1567
electric network. Fig. 4 shows the picture of probe connected to the acquisition card during a test phase in the laboratory. Voltage sensors are dividers able to partition the input voltage to preserve the analog input stage of the
Rome
acquisition card channel. As required by the rule, the maximum input voltage is 6,000 V. As shown in Fig. 5 Palermo
Fig. 1 Geographical position of Rome and Palermo.
The probe is based on an Advantech Industrial Personal Computer connected to an ad hoc acquisition card [2-4] developed by us. This has eight channels, four connected to voltage sensors and four connected to current ones to pick up voltages and currents of three phases (R, S, T) and the neutral N wire, of a common
the voltage sensors have two linear in-out characteristics: the first (normal operation) has a reduction ratio of 1:100, while the other has a reduction ratio of 1:12,000. To assure the higher accuracy in the current measurements, the chosen sensors are Rogowski coils. These, compared with others as the amperometric clamps or the current transformers, assure that the harmonic content of the input signal is unchanged. Fig. 6 shows the application scheme of Rogowski coils.
Probes Probes
Rome
Palermo
Fig. 2 Probe distribution in Rome and Palermo.
Fig. 3 Block scheme of the probe.
1568
Geographical monitoring of Electrical Energy Quality determination: The Problems of the Sensors
ADSL modem
Phone controller
Battery
Vu (f)
PC
Battery charger
Acquisition board
Dividers (0.1%) Uniformity Band
Rogowski coil and integrator (1%)
10 Hz
10 kHz
Log (f/10 Hz)
Fig. 7 Uniformity band comparison between voltage divider and Rogowski coil and integrator.
In case of civil suit, one of the most important problems is to determine as more accurately as possible the time of the fault. In fact, it would be desirable that the synchronism of the instrument, and particularly, of the sampling, were linked to the National Time Standard.
Fig. 4 The probe.
Vu Vi Vu
3. Timing Vi Fig. 5 Voltage pick up with its transfer function. The opposite zeners in the circuit also allow to control a possible voltage spike of 6 kV in accordance with the norm.
Fig. 6 Rogowski coil current detector.
They give an output signal proportional to the derivative of the measured current; so the integration of their output is necessary before to pass the signals to the acquisition card. Fig. 7 shows the uniformity band characteristics of the voltage divider, at 0.1% accuracy, and the current sensors, at 1% accuracy (foreseen in case of current and power measurements). The characteristics are perfectly overlapped between 10 Hz to 10 kHz.
The time base of the sampling of the voltage signal must be synchronized in all probes which are spatially also very far. We decided to use the most stable time base present on the Italian territory coming from INRiM (National Italian Metrological Institute), the INRiM, which provides a clock signal with a stability of about 10-10 s directly to the Telecom Italia telephone exchanges. The right identification of anomalies in the mains, in relation to the perceived quality, is strictly joined with the synchronization problem of probes placed in different places. Nowadays, this limits the choice of geographic position of the probes obliging us to put them inside the voltage cabins placed close to the telephone exchanges equipped with synchronism signal. Fig. 8 shows the structure of the acquisition card developed by us and placed inside the probes.
4. Economic Importance of the Application The market is the virtual meeting place on which, products and/or services (goods) are offered to potential interested customers willing to pay a price.
Geographical monitoring of Electrical Energy Quality determination: The Problems of the Sensors
1569
categories with the technical parameters defined by the Voltage dividers in Fig. 5
international norms. So the right price of the perceived quality is joined with the measurement accuracy of the normed parameters. These last are almost all defined from the frequency of the network. For an accurate assessment of the perceived power
Fig. 8 Acquisition board.
The price of a good is agreed upon on the basis of several elements, not least, the quality of the product. This lastis lexically defined as the â&#x20AC;&#x153;level of excellenceâ&#x20AC;?. The ISO (International Organization of Standardization) in the norm 9,000 defines quality as â&#x20AC;&#x153;Whatever the customer perceives good quality to beâ&#x20AC;? [5]. The user should determine the price of the goods also according to his own perception of the quality. Electrical energy is at the same time a good and a service diffused on geographic basis, to which everybody can access, and that should be the same for all. The CEI EN 50160 is the norm that defines all the characteristics of the electricity supplied by public distribution systems [6] defining thirteen technical parameters. Distributors see the electrical energy usersas loads which produce a series of variations on ideal voltage waveform. This is due to the constant growth of non linear loads connected to the electrical network, which deteriorate the quality of the electrical energy. This deterioration varies over the territory and the perception of the electrical quality analyzed at the same POD (point of delivery) can be judged very differently according to the userâ&#x20AC;&#x2122;s needs. The basic idea is that users should pay for electrical energy on the base of an own judgment quality which needs of objective measurements. Regarding the assessment of the electrical energy quality, in literature, the established six quality categories has been studying, in relation to the perception of the users. The results of these investigations shown the correlations of these
quality and to correctly determine the right price to pay, you must have very accurate measurement of the frequency (around a hundredth of Hz). Wanting use the already shown probe to determine a fairer energy pricing, it is necessary to measure the mains frequency with the requested accuracy: The stable time base is not sufficient, it is necessary also to develop suitable measurement algorithms.
5. The Mains Frequency Measurements In our applications, to measure the real pulsation Ď&#x2030; of the mains signal, it has been used the CFA (curve fitting algorithm) deleting the structural recursivity by an original modification [7-11]. This algorithm has a complexity similar to Fourier analysis, but its main advantage is the evaluation speed. The algorithm allows to extract the frequency with an higher precision. In our case the observation window length is T = 0.02 s (this means 500 samples each period of the rated value of the pulsation Ί). Considering the recursive relations: đ??śđ?&#x2018;&#x2013; đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; =
đ?&#x2018;&#x2021; đ?&#x2018;&#x152; 0
đ?&#x2018;&#x2020;đ?&#x2018;&#x2013; đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; =
đ?&#x2018;&#x2021; đ?&#x2018;&#x152; 0
đ?&#x2018;Ą đ?&#x2018;Ą đ?&#x2018;&#x2013; cos đ?&#x153;&#x201D;đ?&#x2018;Ą dđ?&#x2018;Ą ; đ?&#x2018;&#x2013; = 0, 1, â&#x20AC;Ś (1) đ?&#x2018;Ą đ?&#x2018;Ą đ?&#x2018;&#x2013; sin đ?&#x153;&#x201D;đ?&#x2018;Ą dđ?&#x2018;Ą ; đ?&#x2018;&#x2013; = 0, 1, â&#x20AC;Ś(2)
The proposed version of the CFA algorithm permits to evaluate the difference between Ď&#x2030; from the ideal grid pulsation (50 Hz â&#x;ś â&#x201E;Ś = 100Ď&#x20AC;), solving a polynomial equation, avoiding recursion: đ?&#x2018;&#x201D; đ?&#x2018;&#x2013; = 0 đ?&#x2018;&#x17D;đ?&#x2018;&#x2013;
δĎ&#x2030;đ?&#x2018;&#x2013; = 0 where, the coefficients ai can be calculated as: đ?&#x2018;&#x17D;đ?&#x2018;&#x2013; =
2đ?&#x153;&#x201D; đ?&#x153;&#x2022; i đ??ˇđ?&#x2018;&#x2019;đ?&#x2018;&#x203A; đ?&#x2018;&#x2013;!
đ?&#x153;&#x2022;đ?&#x153;&#x201D; i
+
2
đ?&#x153;&#x2022; iâ&#x2C6;&#x2019;1 đ??ˇđ?&#x2018;&#x2019;đ?&#x2018;&#x203A;
iâ&#x2C6;&#x2019;1 ! đ?&#x153;&#x2022;đ?&#x153;&#x201D; iâ&#x2C6;&#x2019;1
Num and Den are defined as:
(3)
1 đ?&#x153;&#x2022; đ?&#x2018;&#x2013; đ?&#x2018; đ?&#x2018;˘đ?&#x2018;&#x161;
â&#x2C6;&#x2019; i!
đ?&#x153;&#x2022;đ?&#x153;&#x201D; đ?&#x2018;&#x2013;
đ?&#x153;&#x201D; =Ί
(4)
1570
Geographical monitoring of Electrical Energy Quality determination: The Problems of the Sensors
đ?&#x2018; đ?&#x2018;˘đ?&#x2018;&#x161; = đ??ś02 đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; + đ?&#x2018;&#x2020;02 đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; (5) đ??ˇđ?&#x2018;&#x2019;đ?&#x2018;&#x203A; = đ??ś1 đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; đ?&#x2018;&#x2020;0 đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; â&#x2C6;&#x2019; đ?&#x2018;&#x2020;1 đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; đ??ś0 đ?&#x153;&#x201D;, đ?&#x2018;&#x2021; (6) The frequency measurement has been realized on the direct sequence Vn of the tri-phase obtained thanks to the Fortescue decomposition made using the transformation: đ?&#x2018;&#x2030;đ?&#x2018;§ 1 1 đ?&#x2018;&#x2030;đ?&#x2018;? = 1 1 đ?&#x203A;ź 3 đ?&#x2018;&#x2030;đ?&#x2018;&#x203A; 1 đ?&#x203A;ź2
1 đ?&#x203A;ź2 đ?&#x203A;ź
đ?&#x2018;&#x2030;đ?&#x2018;&#x2026; đ?&#x2018;&#x2030;đ?&#x2018;&#x2020; đ?&#x2018;&#x2030;đ?&#x2018;&#x2021;
(7)
with 1
đ?&#x203A;ź = â&#x2C6;&#x2019;2 +đ?&#x2018;&#x2014; 3
(8)
Normally, the zero sequence Vz and the negative sequence Vn are null. Sequence component analysis plays an essential role in analyzing power system faults and explains some power system phenomena. The measurement method uses the integration of the noise in each measurement points: This strongly reduces the mains noise. The method strongly depends by the accuracy of the single terms of the relations (1â&#x20AC;&#x2122;) e (1â&#x20AC;&#x2122;â&#x20AC;&#x2122;), developed up to i = 4, for g = 2 in Eq. (3). The accuracy of δĎ&#x2030; = â&#x201E;Ś - Ď&#x2030; (9) is extremely complex to calculate, but the calculus can be simplified starting from the real hypothesis that the deviations from â&#x201E;Ś are under 1 Hz reaching a relation as: Î&#x201D;δĎ&#x2030; â&#x2030;&#x2C6; (50)2-n (10) where, n is the number of bits of the analogical-digital conversion card. If n = 10 bits, the accuracy is equal to: Î&#x201D;δĎ&#x2030; â&#x2030;&#x2C6; 5 Ă&#x2014; 10-2 rad/s (11) This considerations have led to adopt in Fig. 3, an â&#x20AC;&#x153;Analog Front-end & Sensorsâ&#x20AC;? with an accuracy of 0.1% on a uniformity band of 100 kHz and an â&#x20AC;&#x153;Acquisition
(expressed in percentage of the fundamental) recorded in 2 weeks. The graph shows an anomalous peak in which the harmonic overcomes the 16%. Fig. 10 points out the peak of Fig. 9 showing the three-phase voltage recorded by the proposed system during the peak. The graph clearly shows the high harmonic pollution (especially 11th and 13th). Fig. 11 shows the 17th harmonic trend recorded for each phase. This harmonic is clearly joined with the human activity: Working days present an increasing of the harmonic pollution in the working hours. Saturday and Sunday (16th November and 17th November) this peak is smaller.
7. Future Improvements For commercial aims, a better definition of frequency difference should be necessary. This is achievable with a higher number of bits of acquisition card, but increasing the stability of reference time base.
the not the the
Fig. 9 11th harmonic level of each phase recorded in 2 weeks. The graph shows an anomalous peak.
Boardâ&#x20AC;? with a resolution of 11 bits (Fig. 7).
6. Tests and Measurements In this section, some tests and measurements directly obtained by the probes are shown. The first graph shows the trend of the 11th harmonic of each phase
Fig. 10 Three-phase voltage recorded by the proposed system. This graph shows the high harmonic pollution during the peak shown in Fig. 9.
Geographical monitoring of Electrical Energy Quality determination: The Problems of the Sensors
1571
The f1* measured is in relation with f1 perceived by the following equation: đ?&#x2018;&#x201C;1
đ?&#x2018;&#x201C;1â&#x2C6;&#x2014;
â&#x2030;&#x2026;
đ??ť 2 â&#x2C6;&#x2019;đ??ž đ??ž đ??ťâ&#x2C6;&#x2019;1
đ??ť â&#x2C6;&#x2019;đ??ž
+ đ??ť đ??ť2 â&#x2C6;&#x2019; đ??ž
(14)
8. Conclusions
Fig. 11 17th harmonic level recorded for each phase.
This result could be obtained exploiting the GPS (global positioning system): In this case the definition of the time base would become a problem of sensors. This would make the positioning of the probes on the territory independent by the Telecom synchronism signal allowing a less constrained positioning. We are studying the possibility to connect the probes with the GPS system realizing a new time reference constantly locked to the cesium-derived GPS satellites carriers (L1, L2) by means of a new GPS receiver architecture able to obtain the carriers battement frequency with an accuracy close to the atomic clock [12, 13]. Resolve the problem of the Doppler Effect is necessary to apply the Radovic operator [14]: đ?&#x2018;&#x2018;
đ?&#x2018;? đ??ż1
đ?&#x2018;?
đ?&#x2018;&#x2019; đ?&#x2018;&#x201C;1
đ?&#x2018;?â&#x2C6;&#x2019;đ?&#x2018;Ł đ?&#x2018;Ą
1
đ?&#x2018;&#x2018; đ?&#x2018;&#x201C;1
đ?&#x2018;?
=đ?&#x2018;&#x201C;
(12)
where, c is the light speed, v(t) is instantaneous satellite speed and f1 is the frequency perceived at the earth surface, modified by the Doppler Effect and d is the differential operator. Knowing the defined ratio đ??ż
154
đ??ž = đ??ż2 = 120 1
đ?&#x2018;&#x201C;2 đ?&#x2018;&#x201C;1
=
đ?&#x2018;&#x201C;2 â&#x2C6;&#x2014; đ?&#x2018;&#x201C;1 â&#x2C6;&#x2014;
References [1]
[2]
(13)
and being đ??ť=
This work faces the fundamental problem of sensors for the PQ monitoring on the territory, analyzed joining the information coming from the technical PQ parameters measured by a probe developed by us, with the concept of â&#x20AC;&#x153;perceived qualityâ&#x20AC;?. It is long time that, the international community is trying to do to accept the idea to pay the good â&#x20AC;&#x153;electrical energyâ&#x20AC;? on the base also of its quality. Now the intrinsic objective difficulties related to the definition of this concept are starting to be faced and correlated to the technical parameters scientifically defined. When a customer has to pay for a good or a service, they want the maximum accuracy. In the presented work, it was highlighted as this objective can be essentially faced working on the accuracy characteristics of the sensors used. It has also identified a new application belonging to the electromagnetic sensors necessary for a better definition of the time base starting from the GPS.
[3]
(14)
the ratio verified on the earth after 77 (least common multiple) periods of f1 (measured f1*) that can be calculated by a phase measuring the f2 (measured f2*) at the frequencies of GHz, utilizing the Wang, Mao, Liu [15] procedure.
[4]
Henryk Markiewicz & Antoni Klajn, Voltage characteristics of electricity supplied by public distribution system, 2009. M. Caciotta, F. Leccese, A. Trifiro, The perceived power quality of electrical energy: An assessment in Italy, in: Proceedings of the XVII IMEKO World Congress, Metrology for a Sustainable Progress, Rio de Janeiro, Brazil, Sep. 17-22, 2006. M. Caciotta, S. Giarnetti, F. Leccese, D. Trinca, A multi-platform data acquisition device for power quality metrological certification, in: Proc. of the 9th International Conference on Environment and Electrical Engineering, Prague, Czech Republic, May. 16-19, 2010. M. Caciotta, S. Giarnetti, G.L. Cinquegrani, F. Leccese, D.Trinca, Development and characterization of a multi-platform data acquisition system for power quality metrological certification, in: Proc. of International
1572
Geographical monitoring of Electrical Energy Quality determination: The Problems of the Sensors
Conference on Renewable Energies and Power Quality, Las Palmas de Gran Canaria, Spain, Apr. 13-15, 2011. [5] GPS10 RB-10 MHz Disciplined Rubidium Frequency Standards Brochure, Precision Test Systems Ltd., 1997. [6] UNI EN ISO 9000, Quality Management Systems, International Organization for Standardization, 2009. [7] F. Leccese, Rome: A first example of perceived power quality of electrical energy, in: Proc. of the Seventh IASTED International Conference on Power and Energy Systems—EuroPES 2007, Palma de Mallorca, Spain, Aug. 29-31, 2007. [8] F. Leccese, Rome, A first example of perceived power quality of electrical energy: The telecommunication point of view, in: Proc. of International Telecommunications Energy Conference, Rome, Italy, Sep. 30-Oct. 04, 2007. [9] F. Leccese, The perceived power quality way as new frontiers of relationship between customers and producers, in: Proc. of the 7th EEEIC International Workshop on Environment and Electrical Engineering, Wroclaw-Cottbus, Poland, May 5-11, 2008. [10] F. Leccese, Analysis of power quality data on some telecommunication sites in Rome, in: Proc. of The Eighth IASTED International Conference on Power and Energy Systems—EuroPES 2008, Corfù, Greece, Jun. 23-25,
2008. [11] M. Caciotta, F. Leccese, E. Piuzzi, Study of a new gps-carriers based time reference with high instantaneous accuracy, in: Proc. of International Conference on Electronic Measurement & Instruments, Xi’An, China, Aug. 16-18, 2007. [12] M. Caciotta, F. Leccese, S. Pisa, E. Piuzzi, A voltage controller oscillator for obtaining a frequency reference constantly locked to L1 GPS carrier for power quality assessment applications, in: Proc. of the 16th IMEKO TC4 Symposium Exploring New Frontiers of Instrumentation and Methods for Electrical and Electronic Measurements, Florence, Italy, Sep. 22-24, 2008. [13] M. Caciotta, A.D. Liberto, F. Leccese, S. Pisa, E. Piuzzi, A system for obtaining a frequency reference constantly locked to L1 GPS carrier for distributed power quality assessment, in: Proc. of the International Conference on Electronic Measurement & Instruments, Beijing, China, Aug. 16-18, 2009. [14] A. Radovic, Frequency shift that is caused by observer movement, Journal of Theoretic 5-4 (2003). [15] Z. Wang, L. Mao, R. Liu, High-Accuracy amplitude and phase measurement for low-level RF systems, IEEE Transaction on Instrumentation and Measurement 61 (2012) 912-921.
Journal of Communication and Computer 10 (2013) 1573-1576
Controller Design for Networked Control System with Uncertain Parameters Caixia Guo, Xuebing Wu and Xinwei Yang College of Physics and Electrical Engineering, Henan Normal University, Xinxiang 453007, China
Received: November 04, 2013 / Accepted: December 05, 2013 / Published: December 31, 2013. Abstract: This paper deals with the problem of controller design for NCSs (networked control systems) with uncertain parameters and network induced time-delays. In the case of long time delay constraint and data package loss, the closed-loop model of a class of NCSs is presented for controlled objects with uncertain parameters. Based on Lyapunov stability theory combined with LMIs (linear matrix inequalities) techniques, the design method of state feedback controller for NCSs with asymptotic stability is given in the form of LMIs. Numerical example and simulation suggest that the results are effective and are an improvement over previous ones. Key words: Network induced time-delays, linear matrix inequality, networked control systems.
1. Introductionď&#x20AC; Closed-loop NCSs (network control systems) are constructed via network information transmission for sensors, controllers and actuators located in different position. The introduction of real-time communication network makes the analysis and design of closed-loop NCSs more difficult [1-4]. The network induced time-delays, which either constant or time-varying, can degrade the performance of NCSs and can even destabilize the system. Since NCSs have infinite dimensionality, it is difficult to apply previous well-known stability theory. Also, in practice, the plant almost present some uncertainties because it is very difficult to obtain an exact mathematical model due to environmental noise, uncertain or slowly varying parameters, etc. [5-6]. However, the stability analysis and design of MIMO NCSs became more challenging due to the complexity of long time delay and uncertainty of delay discrete model. System modelling and stability Corresponding author: Caixia Guo, master degree, lecturer, research fields: networked control systems, robust control system. E-mail: 021106@htu.cn.
investigation is the basis to analyze the performance of MIMO NCSs and the design of controllers. Some research has been done [3, 4, 7-10], but most attentions are focused on NCSs with certain model and the stability analysis which is conducted for short time delay. J.H. Zhang [7] established a NCS for the controlled object with uncertain parameters based on a flexible feedback controller and gave the allowable upper bound of time delay for closed-loop system while the discrete system model and package loss were not considered. In this paper, we extend the results of Ref. [7] to uncertain NCSs, and propose the method of controller design in terms of LMIs (linear matrix inequalities). Therefore, our results are more general than the existing ones. The rest of the paper is organized as follows. In Section 2, the model of NCSs to be studied is formulated and related preliminaries are presented. In Section 3, the main results are given. In Section 4, numerical examples and simulation are presented to illustrate our proposed results are effective. Finally, the conclusions are given in Section 5.
1574
Controller Design for Networked Control System with Uncertain Parameters
2. Modeling of Networked Control System Considering the uncertainty of the actual model parameter, the controlled object can be described in discrete state as: x k 1 A A x k B B u k y k C C x k
(1)
investigated. Lemma 1 [9]. For a positive integer d i ，if there exists positive matrices and P, Q, Z , X X 11 X 21
positive integer, the state feedback controller of the controlled object as given in Eq. (1) is (2) uk Kxk d k As package loss may happen when the sensor sends data package, switch matrix F is introduced and placed between input matrix and feedback gain matrix as:
N1 , N 2 are matrix of
arbitrary appropriate dimensions, following matrix inequality hold:
11 12 T 22 12 H A I HA 1 K
n where, xk R is the state of the controlled object; yk R m is the output of the controlled object; uk R l is the input of the controller; A, B, C
stand for uncertain state matrix, uncertain input matrix and uncertain output matrix, respectively. Assuming the total time delay of network transmission is bounded, and d1 d k sc ca d 2 is satisfied. Where, d i is a
X 12 0, also X 22
X 11 X T 12 N 1T
X 12 X 22 T2T
making
the
A1 I T H AKT H H
0
N1 N 2 0 Z
(5)
(6)
where,
11 P A1 I A1 I P T
N1 N1T d 2 d1 1 Q d 2 X11 12 PAK N1 N 2T d 2 X12 22 N 2 N 2T Q d 2 X 22 H P d2 Z
Then, the system expressed in Eq. (4) is F diag f1 k , ... f m k mm asymptotically stable, if time delay bounded condition (3) is satisfied. where, Theorem 1. For a positive integer d i ，if there 1 Sensor node i data transmission successful exists positive matrices and D, E , L , fi (k ) , im 0 Sensor node i data transmission failure Y Y sion successful Y 11 12 0 , also, G1 , G2 , U are matrices of , im Y21 Y22 ission failure The discrete-time model of MIMO NCSs is: arbitrary appropriate dimensions and 0 , making x k 1 A A x k B B KFx k d k the following matrix inequality established: (4) 1 x k AK x k d K
11 T 12 T A1 I D
In the condition of package loss, the design objective of the MIMO NCS is to obtain k value of the controller and makes the whole system asymptotically stable in case of time delay bounded.
Y11 Y T 12 G1T
3. Design of the Controller The stability criterion for the system expressed in Eq. (4) is given in this part and the controller design of closed-loop control network system is also
where,
12 22 B1U
Y12 Y22 T 2
G
D A1 I 0 (7) U T B1T T
D 1 d 2
G1 G2 0 1D
(8)
Controller Design for Networked Control System with Uncertain Parameters
Y11 Y T 12 G1T
A I D D A I T 1 1 11 G1 G1T d 2 d1 1 E d 2Y11 12 B1U G1 G2T d 2Y12 T 22 G2 G2 E d 2Y22
G2 P 1 N 2 P 1 , U KFP 1 , L Z 1 ,
P 1 X 12 P 1 . P 1 X 22 P 1
T D A1 I
22
0 L D d 2 U T B1T
B1U
G
4. Simulation Results For discrete system with uncertain model: xk 1 A Axk B Buk 0.0002 0.0244 0.0122 A , B 0.0122 , 0.0122 0.8089 0.0001 0 0.002 A , B 0 0 0 . 1
(9)
linear matrix inequalities of Eq. (7) and Eq. (8) are solved assuming F diag 1, 0 . The solution is K 0.2998 7.483 . Fig. 1 shows the simulation result. It can be seen from Figs. 1 and 2 that state vector x1 , x2 is quickly stabilized. Therefore, NCS is asymptotically stable under the influence of the controller. x1
2.5
state response-x1
2
1.5
1
0.5
0
1
Fig. 1 Response curve of state vector x1.
2
(11)
where,
3
-0.5 0
(10)
Suppose the sampling period is T = 0.1 s and time delay is 5T d K 10T . In case of package loss,
then, it can be calculated as 12
T 2
matrix of the controller is K UDF 1 .
D P 1 , E P 1QP 1 , G1 P 1 N1P 1 ,
11 T 12 A1 I DT
Y22
where, DL1 D in Eq. (10) is nonlinear. If L D , Eq. (7) and Eq. (8) can be obtained. The parameter
then, the system with uncertain parameters as expressed in Eq. (4) is asymptotically stable, if time delay bounded condition is satisfied. The above theorem can be proved by the following deduction. Eq. (5) is left and right multiplied by diag P1 , P1 , H 1 and Eq. (6) is left and right multiplied by diag P1 , P1 , P1 , supposing
P 1 X 11 P 1 Y 1 T 1 P X 12 P
G1 G2 0 DL1D
Y12
1575
3
4
5 time/s
6
7
8
9
10
Controller Design for Networked Control System with Uncertain Parameters
1576
2 x2
x2-state response
1.5
1
0.5
0
-0.5
-1
0
1
2
3
4
5 time/s
6
7
8
9
10
Fig. 2 Response curve of state vector x2.
5. Conclusions In the paper, a discrete model for long time delay bounded MIMO NCS is established and the stability analysis is conducted in this paper for controlled objects with uncertain model. The observer and the controller are designed by means of separation principle and the LMIs form for asymptotic stable closed-loop system is also given. Therefore, our results are more general than the existing ones.
References [1]
[2]
[3]
F.L. Lian, W. Moyne, D. Tilbury, Optimal controller design and evaluation for a class of networked control systems with distributed constant delays, in: Proceeding of the American Control Conference, Alaska, USA, 2002, pp. 3009-3014. W.H. Fan, Modeling and control of network control system, Nanjing University of Science and Technology, Nanjing, China, 2004. C.X. Guo, Z.R. Xiang. Controller design of MIMO network control systems, Control Engineering of China
14 (2007) 263-265. H. Chan, U. Ozguner, Optimal control of systems over a communication network with queues via a jump system approach, in: Proceeding of the 4th IEEE Conference on Control Applications, Albany, NY, USA, 1995, pp. 1148-1153. [5] B. Liu, Y.Q. Xia, S.M. Magdi, et al., New predictive control scheme for networked control systems, Circuits Syst Signal Process 31 (2012) 945-960. [6] Y.Y. Hou, Reliable stabilization of networked control systems with actuator faults, Nonlinear Dyn 71 (2013) 447-455. [7] J.H. Zhang, Y. Chen, S.C. Wang, Stability analysis of networked control system with resilient controller, Measurement & Control Technology 29 (2010) 91-92. [8] L. Yu, Robust Control-Approaches for Linear Matrix Inequality, Qinghua University Press, Beijing, 2002. [9] Y. He, Modeling and control of network control system, delay-dependent robust stability and stabilization based on free-weighting matrices, Central South University, Changsha, China, 2004. [10] H.C. Yan, X.H. Huang, M. Wang, H. Zhang, New delay-dependent stability criteria of uncertain linear systems with multiple time-varying state delays, Chaos, Solitons and Fractals 37 (2008) 157-165. [4]