BERIN KOCABAS FELIPE OEYEN
NATURAL SYSTEMS & BIOMIMETICS SOCIAL INSECTS & ALGORITHMS
Natural Systems & Biomimetics Emergent Technologies & Design Studio Directors: Dr. Michael Weinstock and Dr. Elif Erdine. Studio Tutors: Alican Sungur, Abhinav Chaudhary, and Eleana Polychronaki
1
ABSTRACT
The following report depicts the development of an agent-based system running in Grasshopper for Rhino, with Python programming language and Culebra plugin for agent simulation, as part of the Architectural Association Emtech Biomimetics Seminar. The agents studied in order to extract a particular behaviour were Ants in general and Lassius Niger Ants in particular. In the Research Stage we extract a number of behaviours and characteristics found in ants’ research conducted by biologists and computer scientists, in order to abstract them as parameters in an agent-based system. The Stage 1 section of the report shows the initial system simulations ran, changing variables through iterations, with the aim of understanding our system and the impact of parameters’ variation in it. In the Stage 2, we refine our system, making it more dynamic and increasing its complexity. We describe the experiments conducted and evaluate the results. Stage 3 consists of the structural and environmental performance analysis of three different outcomes generated by our simulations of Stages 1 and 2, as well as of an interpretation of the obtained results. For Stage 4 we take one of the three outcomes analysed in Stage 3 and further abstract it in order to rationalize it, considering it would be further developed as a pavilion design. Finally we extract conclusions from our system and its outcomes and analyse the possibilities provided by Biomimetics research and agent-based systems.
3
CONTENT
RESEARCH STAGE / SYSTEM LOGIC STAGE 1 / INITIAL SIMULATION TRIALS STAGE 2 / FURTHER SIMULATIONS /ADVANCED TRIALS STAGE 3 / ANALYSIS (WIND/ STRUCTURAL/ SOLAR) STAGE 4 / FINAL STAGE/ PAVILLION
INTRODUCTION The aim of this investigation was to programme an agent-based system for design development purposes. This ought to be done through the study of a given social insect colony and the extraction of a particular behaviour that could be abstracted into a list of actions and parameters. One of the main concepts studied in insect behaviour and essential for our research is the concept of stigmergy (Fig. 1), as defined by Heylighen1 :
“a fundamental mechanism of self-organization: it allows global, coordinated activity to emerge out of local, independent actions.”
Fig. 1: The Stigmergic Feedback Loop. Source: Heylighen F. «Stigmergy as a universal coordination mechanism I: Definition and components.»1
After reading a series of papers describing stigmergy in ants, termites, wasps, and bees, we chose to work with the first ones. We were particularly interested in their ability to carry out coordinated tasks in the movement of objects, as we thought that could be useful for material collection in the design and construction of a new building. After focusing on ants, we found the Lassius Niger Ants’ “cemetery formation”2 behaviour especially intriguing and decided to use it as the starting point in the writing of our computational design system. The purpose of developing this system was to get a different set of outcomes by defining a set of basic inputs and rules, and running a number of iterations with that logic. Having agents generating different outcomes would allow us to compare them, in order to get a design that would satisfy our requirements in terms of space definition, structural stability and environmental performance.
5
6
RESEARCH During our Research Stage we became particularly interested in ants. And more especifically in one particular behaviour, which is called “cemetery formation” (Fig. 2), as described by Dorigo M., Bonabeau E. and Theraulaz G.2 for experiments conducted on Lassius Niger ants and computational simulations (Fig. 3 & 4) of that behaviour: “If dead bodies, or more precisely items belonging to dead bodies, are randomly distributed in space at the beginning of the experiment, the workers will form clusters within a few hours. If the experimental arena is not sufficiently large, or if it contains spatial heterogeneities, the clusters will be formed along the borders of the arena or more generally along the heterogeneities. The basic mechanism underlying this type of aggregation phenomenon is an attraction between dead items mediated by the ant workers: small clusters of items grow by attracting workers to deposit more items. It is this positive feedback that leads to the formation of larger and larger clusters.”
Fig. 2: Real ants clustering behaviour intial state, after 3, 6 and 36 hours. Source: Dorigo, Bonabeau, Theraulaz. «Ant Algorithms and Stigmergy».2
We decided to take the cemetery formation behaviour and abstract it for the generation of our computational system. We also chose to work with ants with different speeds. This is a feature introduced by Lumer and Faieta in their experiment, which is also described in Dorigo, Bonabeau and Tharaulaz2 paper:
Fig. 3: Computer Simulation of the clustering model. Source: Dorigo, Bonabeau, Theraulaz. «Ant Algorithms and Stigmergy».2
“Fast moving ants are not as selective as slow ants in their estimation of the average similarity of an object to its neighbors. The diversity of ants allows to form the clusters over various scales simultaneously: fast ants form coarse clusters on large scales, i.e., drop items approximately in the right coarsegrained region, while slow ants take over at smaller scales by placing objects with more accuracy.” Fig. 4: Simulation of the clustering alogrithm with 10 ants. Source: Dorigo, Bonabeau, Theraulaz. «Ant Algorithms and Stigmergy».2
RESEARCH
7
Another aspect we found interesting was the ants need to remove the corpses from the nest in order to increase the survival rates of the colony in general and of the larvae in particular (Fig. 6). This behaviour is called necrophoresis and is described by Diez L., Lejeune P. and Detrain C.3:
“In order to quantify the influence of necrophoresis on ants, we compared the survival of M. rubra ants in colonies that were limited in their ability to remove corpses (limited removal: LR colonies, N = 15) to that of control colonies which were able to remove them normally (free removal: FR colonies, N = 15).
[…]
Fig. 5: Snapshots of the cemetery formation in a periodic space with a U-shaped inhomogeneity. Source: Martin, Chopard, Albuquerque. «Formation of an ant cemetery: swarm intelligence or statistical accident?».4
We quantified the impact of corpse removal on the demography of ant colonies. Overall, workers in LR colonies survived less than those in FR colonies.
[…]
In order to understand at what time these differences became significant, we compared the survival rate day by day, finding workers’ survival in FR colonies was higher than workers’ survival in LR colonies from Day 8 onwards.
[…] In FR colonies, corpses were removed rapidly, and none remained in the nest after 4 days. In LR colonies, most corpses remained in the nest until the fourth day. After 8 days, workers managed to cut some corpses into pieces and thus succeeded in removing these body parts out of the ‘small-holes’ nest entrance.”
Fig. 6: Relative distance of corpses from the larvae patch. Source: Diez, Lejeune, Detrain. «Keep the Nest Clean: Survival Advantages of Corpse Removal in Ants».3
The last behaviour we found interesting in the dead bodies collection by ants, is that living ants release chemicals that allow other ants to sense they are alive. When they die these chemicals disappear, making it possible for worker ants to pick their bodies (Fig. 7), this creates a positive feedback for the collection of corpses. This behaviour is decscribed by Choe D., Millar J. G., and Rust M. K.5 :
“Our study showed that the chemical stimuli that elicit necrophoric behavior by L. humile workers are continually present, but additional chemicals associated with living workers inhibit this behavior. Shortly after death, these inhibitory chemicals dissipate or are degraded, allowing the latent necrophoric behavior to be triggered.”
Fig. 7: Responses of Argentine ant workers to dead ants and inanimate objects. Source: Choe, Millar, Rust. «Chemical
Signals Associated with life inhibit necrophoresis in Argentine ants».5
8
SYSTEM LOGIC With these behaviours in mind we set our system logic. The agents in our system would behave the same way that the Lasius Niger Ants do while collecting dead body parts. The cemetery formation was abstracted as the grouping of objects into clusters. The removal of bodies from the nest was translated for our system into a movement of our agents away from the center. The creation of a single big cluster described in the read papers at the end of the experiments, would be adapted as a progressive reduction of the distance between our attraction points, eventually behaving as a single attraction point at the end of each simulation. The existence of Fast and Slow ants would be used to have Fast agents appearing throughout a bigger bounding box than Slow agents. Our system would then be structured the following way. A central repeller point would increase its value through iterations. Around it we would set a number of attraction points that would move within a given range and would eventually get closer together. The whole system would be populated with two types of agents with different speeds and moving at different scales. The fastest agents would appear in a bigger bounding box and the slowest agents would appear inside a smaller bounding box. These premises were the beginning of our investigation, and would remain constant as we continued our experiments.
ANTS AS AGENTS Lasius Niger Ant Behaviour Algorithmic Abstraction Parameter Criteria We were now ready to set our first simulations. To do so, we worked in “Grasshopper” for “Rhino”, with “Python” programming language and “Culebra” plugin for agent simulation. To model the agents’ movements, we worked with Cohesion and Separation (Fig. 8 & 9) as defined by Reynolds C.W.6: “Separation steering behavior gives a character the ability to maintain a certain separation distance from others nearby. This can be used to prevent characters from crowding together. [...] Cohesion steering behavior gives a character the ability to cohere with (approach and form a group with) other nearby characters.”
Fig. 8: Cohesion. Source: Reynolds «Steering
Behaviors For Autonomous Characters».6
Fig.9: Separation. Source: Reynolds «Steering Behaviors For Autonomous Characters».6
STAGE 1: INITIAL SIMULATION TRIALS W/ SINGLE AND DOUBLE POINTS The initial setting of the simulations consists of 2 bounding boxes, one inside of the other one, a bigger outer box of 3m x 3m x 1m, from now on referred to as outer boundary, and an inner, smaller one of 2m x 2m x 0.8m called inner boundary. The inner boundary limits the movement of the slow agents, while the outer boundary stands for the fast agents’ movement. The repellent point is located in the bottom center of the larger bounding box, and the attractor points are placed randomly within the range of the smaller box as shown in Figure 10. The attractor points and the repellent point each have their own threshold values that respectively determine the factor of attraction and repulsion. The bigger these threshold values are, the more they affect the agents’ movements.
Attraction Points (Threshold Radius) Fast agents Slow agents Fast Agent Bounding Box Slow Agent Bounding Box
Repellent Point (Threshold Radius) Fig.10: Initial simulation setup
9
10
1: 60 Slow/ 120 Fast Agents /Constant Single Attractor Point Slow Agents: Initial Speed: 0.50 Max Speed: 0.88 Max Force: 2.24 Fast Agents: Initial Speed: 0.50 Max Speed: 3.43 Max Force: 4.27
In simulation 1, 60 slow and 120 Fast agents are observed within a given dimension of bounding boxes: Slow agents were moving in a 2m x 2m x 0.8m while the Fast agents had a bigger boundary of 3m x 3m x 1m. At the end of the iteration, the agents clustered around the repulsion point, as well as along some random paths within the boundary (Fig. 11 & 12). The reason why they also clustered randomly was the stigmergy behaviour of the agents. The overall movement of the agents is observed to be both towards an attraction point and away from the repulsion point, this way creating bundles of movement paths, driven by stigmery.
Stigmergy: View Angle: 116 Cohesion Mag: 2.4 Cohesion Range: 80.6 Seperation Mag: 0.0 Seperation Range: 5.0 Repulsion Force: Threshold Value: 50 Repel Value: 1 Attraction Force : Threshold Value: 50 Attraction Value: 10
Fig.11: Simulation 1, Top view
Fig.12: Simulation 1, Side view
2: 60 Slow/ 120 Fast Agents /Double Stable Attractor Points Slow Agents: Initial Speed: 0.50 Max Speed: 0.88 Max Force: 2.24
In simulation 2, 60 Slow and 120 Fast agents are observed within the same bounding box of Simulation 1. The agents’ movements showed similarity with Simulation 2 (Fig. 13 & 14), as the agents were controlled by the stigmergy, repulsion and attraction forces as well.
Fast Agents: Initial Speed: 0.50 Max Speed: 3.43 Max Force: 4.27
However, in Simulation 2, the number of attraction points is increased to 2, that altered the clustering and the final position of the agents. In Simulation 1, the velocity of the agents decreased and eventually stopped at random locations and at the attraction point. While in Simulation 2, agents mainly stopped at the 2 attraction points but hardly did so at the random locations (there were some agents at random bundling locations, but the number of them was extremely low compared to the rest).
Stigmergy: View Angle: 116 Cohesion Mag: 2.4 Cohesion Range: 80.6 Seperation Mag: 0.0 Seperation Range: 5.0
This final settling of the agents showed that, within a given range, if the attraction points’ threshold value isn’t big enough (50), they are not able to attract the total of 180 agents (as in Simulation 1). Whereas, when the number of attraction points increased, the total threshold value (150) to attract the agents inside the bounding box also increases, and this results in the attraction of the majority of the agents.
Repulsion Force: Threshold Value: 50 Repel Value: 1 Attraction Force 1: Threshold Value: 50 Attraction Value: 10 Attraction Force 2: Threshold Value: 100 Attraction Value: 10
Fig.13: Simulation 2, Top view
Fig.14: Simulation 2, Side view
11
12
2: 60 Slow/ 120 Fast Agents /1 Randomly Moving Attractor Point Slow Agents: Initial Speed: 0.50 Max Speed: 0.88 Max Force: 2.24 Fast Agents: Initial Speed: 0.50 Max Speed: 3.43 Max Force: 4.27 Stigmergy: View Angle: 116 Cohesion Mag: 2.4 Cohesion Range: 80.6 Seperation Mag: 0.0 Seperation Range: 5.0
In Simulation 3, the agent’s number and the bounding boxes remained the same as for Simulations 1 and 2. The difference for this simulation was the random movement of the attractor point. Here, we chose to have a single moving attractor point, to observe the behaviour of the agents with this change. We decided to increase the number of points for the following experiments, in order to see how that affected the system. The behaviour of the agents showed similarity to that of Simulation 1 (Fig. 15 & 16), as they both had a single attraction point. However, the movement of the attractor point in vertical direction, leaded the majority of the agents to move upwards and changed the dynamics of the overall movement path.
Repulsion Force: Threshold Value: 50 Repel Value: 1 Attraction Force/ Random Movement : Threshold Value: 50 Attraction Value: 10
Fig.15: Simulation 3, Top view
Fig.16: Simulation 3, Side view
STAGE 1 RESULTS: INITIAL TRIALS/ 3 SIMULATIONS The objective of the 3 initial simulations conducted on Stage 1, was to understand the effects of repulsion, attraction and stigmergy forces and how the Fast and Slow agents behaved under these (Fig. 17). With these first results, it was decided to continue exploring the movement of the attractor point(s) choosing the programming of a controled movement over a random one. In the final representations of the agents’ movement paths, the purple lines represent the Slow agents, and the white lines stand for the Fast ones. As the bounding box of the Slow agents was smaller than the one of the Fast agents, and the number of Slow agents was also lower, the purple lines appeared in a tighter area and were less dense than the white ones.
1: Single constant attractor point 2: Double attractor constant point 3: Single random moving attractor point
Fig.17: Comparison of Simulations 1, 2 and 3. Perspective view
13
14
STAGE 2: DYNAMIC AND MULTIPLE ATTRACTOR POINTS For Stage 2 we kept the same system logic than for Stage 1, defining the simulation setup with the repeller point in the middle, surrounded by attractor points in the perimeter. Through iterations the repeller point in the middle grows and the attractor points shift their position and move up in the Z direction. As the repeller force grows, the attraction force of the attraction points diminishes. We ran a number of simulations, changing the values for the number of Fast and Slow agents and the relation ratio between them, as well as changing the number of attraction points.
Our code defines one repeller point in the middle and a number of attractor points around it.
Through iterations the repeller point in the middle grows and the attractor points around it shift their position and move in the z direction.
As the repeller force grows, the attraction force of the attraction points diminishes.
Fig.18: System Movement Diagram
15
Fig.19: System Stages overlapped Diagram
16
4: 150 Fast/ 30 Slow Agents /4 Attractor Points Nยบ of Attractor Points 4 Initial distance between points: 3000 mm Final distance between points: 50 mm Nยบ of Repeller Points 1. Threshold Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Repeller Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Nยบ of Fast Agents: 150 Initial Speed: 0.50 Max Speed: 3.43 Max Separation: 15 Threshold Values through iterations: 125, 125, 115, 105, 95, 85, 75, 65, 55, 45, 35, 25 Attraction Values through iterations: 150, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60, 50
In Simulation 4, we increased the number of attraction points to 4. We also increased the value of the Z dimension for the bounding boxes of both Fast and Slow agents, reaching a final height of 4 m. Another difference between this simulation and the former ones was the number and proportion of Fast and Slow agents. We chose to work with a bigger number of Fast agents and a smaller number of Slow ones than before. This results in a high visibility of purple trails and a very low appearance of white ones. Working with 4 attraction points, we can see that the resulting form resembles a square-based pyramid. Another interesting result of the simulation is that by the end, once the height has reached its limit, the agents are being repelled away from the model, due to the increase of the repeller point value and threshold.
Stigmergy Values: View Angle: 116. Cohesion Mag: 2.4. Cohesion Range 80.6. Separation Mag: 0.0. Separation Range 5.0 Nยบ of Slow Agents: 30 Initial Speed: 0.50 Max Speed: 0.88 Threshold Values through iterations: 150, 135, 125, 115, 105, 85, 75, 50, 30, 15, 10, 5 Attraction Values through iterations: 160, 160, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60
Fig.20: Simulation 4, Top view
Fig.21: Simulation 4, Side view
5: 150 Fast/ 30 Slow Agents /4 Attractor Points Nº of Attractor Points 6. Initial distance between points: 3000 mm. Final distance between points: 100 mm Nº of Repeller Points 1. Threshold Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250 Repeller Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250 Nº of Fast Agents: 150 Initial Speed: 0.50 Max Speed: 3.43 Max Separation: 15 Threshold Values through iterations: 125, 125, 115, 105, 95, 85, 75, 65, 55, 45, 35, 25 Attraction Values through iterations: 150, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60, 50
Simulation 5, shows a new increase in the number of attractor points, reaching the number of 6. The number of Fast and Slow agents remains the same than the one at Simulation 4. The bounding box height is different here, reaching a value of 2.5 m. Being the final distance between points higher than the one at the previous simulation, we can see the Slow agents’ trails are more visible. The final form is a low hexagon-based pyramid, with equally distributed purple trails all around its perimeter. We can see that the agents are closer to the model, because the final repeller point value and threshold is only half the one in Simulation 4.
Stigmergy Values: View Angle: 116. Cohesion Mag: 2.4. Cohesion Range 80.6. Separation Mag: 0.0. Separation Range 5.0 Nº of Slow Agents: 30 Initial Speed: 0.50 Max Speed: 0.88 Threshold Values through iterations: 150, 135, 125, 115, 105, 85, 75, 50, 30, 15, 10, 5 Attraction Values through iterations: 160, 160, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60
Fig.22: Simulation 5, Top view
Fig.23: Simulation 5, Side view
17
18
6: 60 Fast/ 120 Slow Agents /6 Attractor Points Nº of Attractor Points 6 Initial distance between points: 3000 mm Final distance between points: 75 mm Nº of Repeller Points 1. Threshold Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Repeller Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Nº of Fast Agents: 60 Initial Speed: 0.50 Max Speed: 3.43 Max Separation: 15 Threshold Values through iterations: 125, 125, 115, 105, 95, 85, 75, 65, 55, 45, 35, 25 Attraction Values through iterations: 150, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60, 50 Stigmergy Values: View Angle: 116. Cohesion Mag: 2.4. Cohesion Range 80.6. Separation Mag: 0.0. Separation Range 5.0
For simulation 6, we kept the number of attractor points in 6, but changed the number and proportion of Fast and Slow agents again. For this experiment, we wanted to check the influence of the Slow agents in the final outcome, that’s why they double the fast ones in number. The bounding box height is again 2.5 m. The final distance between points is slightly higher for this simulation than for the previous one. Nevertheless, the Slow agents’ trails are much more visible than before, due to the increase in number of agents. The final form is much more irregular than the two previous ones, mainly because the Slow agents’ threshold value for the final iterations is lower than the one of the Fast agents. This combined with the high repelling values for the end of the simulation results in the repeller force being stronger than the attraction force, pushing the Slow agents further away from the model (Fig. 24 & 25).
Nº of Slow Agents: 120 Initial Speed: 0.50 Max Speed: 0.88 Threshold Values through iterations: 150, 135, 125, 115, 105, 85, 75, 50, 30, 15, 10, 5 Attraction Values through iterations: 160, 160, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60
Fig.24: Simulation 6, Top view
Fig.25: Simulation 6, Side view
7: 90 Fast/ 90 Slow Agents /3 Attractor Points Nº of Attractor Points 3. Initial distance between points: 3000 mm Final distance between points: 50 mm Nº of Repeller Points 1. Threshold Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Repeller Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Nº of Fast Agents: 90 Initial Speed: 0.50 Max Speed: 3.43 Max Separation: 15 Threshold Values through iterations: 125, 125, 115, 105, 95, 85, 75, 65, 55, 45, 35, 25 Attraction Values through iterations: 150, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60, 50
In Simulation 7, we reduced the number of attractor points to 3, and changed the number and proportion of fast and slow agents to 90 each. The goal of this experiment was to see how the outcome changed by using the same amount of Fast and Slow agents. The bounding box height was also set back to 4 m. The final distance between points is the same that the one in Simulation 4. We can see the slow and fast agents’ trails are intertwined being their number of agents the same. The final form is a regularly-shaped spiralled tower, with some agents being repelled on top and to the sides (Fig. 26 & 27).
Stigmergy Values: View Angle: 116 Cohesion Mag: 2.4 Cohesion Range 80.6 Separation Mag: 0.0 Separation Range 5.0 Nº of Slow Agents: 90 Initial Speed: 0.50 Max Speed: 0.88 Threshold Values through iterations: 150, 135, 125, 115, 105, 85, 75, 50, 30, 15, 10, 5 Attraction Values through iterations: 160, 160, 150, 140, 130, 120, 110, 100, 90, 80, 70, 60
Fig.26: Simulation 7, Top view
Fig.27: Simulation 7, Side view
19
20
8: 90 Fast/ 90 Slow Agents /3 Attractor Points Nยบ of Attractor Points 3 Initial distance between points: 1500/ 3000 mm Final distance between points: 50/ 100 mm Nยบ of Repeller Points 1. Threshold Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Repeller Values through iterations: 20, 40, 80, 100, 130, 150, 200, 250, 300, 350, 400, 500 Nยบ of Fast Agents: 50 Initial Speed: 0.50 Max Speed: 3.43 Max Separation: 15 Threshold Values through iterations: 125, 125, 115, 105, 95, 85, 75, 65, 55, 45, 35, 25 Attraction Values through iterations: 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 150
Simulation 8, was our final one. Here we decided to keep the number of attraction points in 3, but chose to modify the distance between them in order to avoid having an equilateral triangle. The number of Fast and Slow agents remained the same than for the previous simulation. The bounding box height also remained the same. By looking at the final outcome, we can appreciate the existence of two main paths. This is mainly due to the irregular distribution of the attraction points. Being two points closer to each other than to the third one, they work as a single attraction point and the third one works as another one. The final form is a double helix with a wide base and a low height (Fig. 28 & 29). Again we can observe some agents being repelled at the highest positions, due to increase of the repeller force by the end of the simulation.
Stigmergy Values: View Angle: 116 Cohesion Mag: 2.4. Cohesion Range 80.6 Separation Mag: 0.0 Separation Range 5.0 Nยบ of Slow Agents: 50 Initial Speed: 0.50 Max Speed: 0.88 Threshold Values through iterations: 150, 135, 125, 115, 105, 85, 75, 50, 30, 15, 10, 5 Attraction Values through iterations: 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 160
Fig.28: Simulation 6, Top view
Fig.29: Simulation 6, Side view
STAGE 2 RESULTS After running eight different simulations, slightly changing some parameters for each, but keeping the same logic for all, we could have a better understanding of our system, its variables, and the impact each one had on the final outcome. We can appreciate the potential of our agent-based system to generate multiple outcomes. The result of each simulation is very different from the others. This divergence is achieved only by changing the number of attraction points, or the proportion of Fast and Slow agents. We took screenshots of every simulation we ran from Top and Side view, in order to choose which geometries we would take further to the Analysis Stage, where we would conduct environmental and structural behaviour tests. We selected two digital models from Stage 2 and one from Stage 1 to compare the results. The choosing criteria was selecting the ones that seemed to generate an interesing inner space and a sheltering, trying to have three contrasting models, that would give us different results in our future analysis. In order to turn the trail lines of our agents into a digital input we could evaluate with the environmental and structural software, we meshed them using “Cocoon” plugin for “Grasshopper”. This software generated geometries made out of a series of pipes with variable section, which were the first translation of our simulations into spatial models that would be analyzed in Stage 3.
21
22
STAGE 2 RESULTS
1
4
Fig.30 Stage 1 and 2 simulation outcomes, Top and Side views
2
5
6
PLAN
7
ELEVATION
ELEVATION
PLAN
23
3
8
24
STAGE 3: ANALYSIS ANALYSIS 1 Simulation 3: 60 Slow/ 120 Fast Agents Single Attractor Point Random Movement
Fig.31 : Simulation 3 meshed geometry. Perspective View
ANALYSIS 2 Simulation 5: 150 Fast/ 30 Slow Agents 6 Attractor Points Movement by Feedback
Fig.32 : Simulation 5 meshed geometry. Perspective View
ANALYSIS 3 Simulation 4: 150 Fast/ 30 Slow Agents 4 Attractor Points Movement by Feedback
Fig.33 : Simulation 4 meshed geometry. Perspective View
25
In Stage 3, the selected simulations are analyzed in detail. By giving them a thickness with “Cocoon”, the agent trails became geometries suitable to analyze structurally and spatially. Wind, structure and solar analysis are applied to these 3 simulations in order to select 1 of them at the end of this stage, to take further to the pavillion design stage. In the wind analysis part, wind pressure and wind velocity applied on the geometry surfaces are depicted as color charts. Moreover, wind vectors are simulated on the geometries and represented through waves that can be observed on the simulations, while the wind pressure simulated is plotted as line graphs. For the structural analysis, “Karamba” is used and the “Displacement”, “Principal Stress 1” and “Principal Stress 2” charts are extracted. As for the solar analysis, several methods are applied. First, the radiation is drawn on the geometry surfaces and graphed as colored geometries. Second, the radiation is reflected on the ground plane of the geometry in order to conduct shadow analysis studies.
26
ANALYSIS 1 As Figure 34 illustrates, maximum wind velocity on the surface of the Analysis 1 is 14.362 m/s and maximum pressure is 67.414 Pa. The velocity and pressure values can be read from the charts as well as they are reflected as range of colors on the top view and the elevations. The Karamba Analysis in Figure 35 shows that the major displacement value is 3.10 cm and the Principle Stresses 1 and 2 are 2.01 kN/cm2 and 7.84 kN/cm2 respectively. In Figure 36, solar analysis is showing the radiation upon the surface during a 10 hours period. From the colored diagram of the surface, it can be seen that the density of the geometry has effects on the inner solar radiation as the surfaces avoid the sunlight from penetrating through the bottom parts.
Fig.34 : Analysis 1/ Wind Analysis
Fig.35 : Analysis 1/ Structural Analysis
Wind speed: 10.000 (m/s) Length: 0.661 (m) Width: 0.467 (m) Height: 0.237 (m) Voxel Size: 0.003 (m)
Fig.36 : Analysis 1/ Solar Analysis Weather File Source: London Wea Ctr St James Park - GBR Month: May Hours: 8:00 am /18:00 pm
Displacement P.Stress 1
P.Stress 2
ANALYSIS 2
27
In Figure 37, maximum wind velocity on the surface of the Analysis 2 is 13.273 m/s and maximum pressure is 76.676 Pa. The Karamba Analysis in Figure 38 shows that the major displacement value is 0.00747 cm (where it was 3.10 cm for Analysis 1) and the Principle Stresses 1 and 2 are 0.161 kN/cm2 (2.01 kN/cm2 in Analysis 1) and 0.0564 kN/cm2 (7.84 kN/cm2 in Analysis 1) respectively. In Figure 39, solar analysis of the coloured diagram of the surface, shows variation of solar radiation density along the perimeter and the inner part.
Fig.37 : Analysis 2/ Wind Analysis
Fig.38 : Analysis 2/ Structural Analysis
Wind speed: 10.000 (m/s) Length: 0.661 (m) Width: 0.467 (m) Height: 0.237 (m) Voxel Size: 0.003 (m)
Fig.39 : Analysis 2/ Solar Analysis Weather File Source: London Wea Ctr St James Park - GBR Month: May Hours: 8:00 am /18:00 pm
Displacement P.Stress 1
P.Stress 2
28
ANALYSIS 3 In Figure 40, maximum wind velocity on the surface of the Analysis 3 is 21.270 m/s ( where the value was 14.362 m/s and 13.273 m/s on the previous analyses and the Analysis 3 show the highest wind velocity) and maximum pressure is 311.667 Pa (it was 67.414 Pa and 76.676 Pa for Analysis 1 & 2 respectively, so the Analysis 3 shows the highest maximum pressure value on the surface). The Karamba Analysis in Figure 41 shows that the major displacement value is 0.0370 cm (where it was 3.10 cm and 0.00747 cm for Analysis 1 & 2 respectively) and the Principle Stresses 1 and 2 are 0.377 kN/cm2 (2.01 kN/cm2 in Analysis 1) and 0.0707 kN/cm2 (7.84 kN/cm2 in Analysis 1) respectively. In Figure 42, solar analysis of the coloured diagram of the surface, shows the variation of solar radiation density at certain parts and as the geometry of Analysis 3 is the highest one among the 3 Analysis, the solar radiation is smaller on the lower parts of the surface.
Fig.40 : Analysis 3/ Wind Analysis
Fig.41 : Analysis 3/ Structural Analysis
Wind speed: 10.000 (m/s) Length: 0.661 (m) Width: 0.467 (m) Height: 0.237 (m) Voxel Size: 0.003 (m)
Fig.42 : Analysis 3/ Solar Analysis Weather File Source: London Wea Ctr St James Park - GBR Month: May Hours: 8:00 am /18:00 pm
Displacement P.Stress 1
P.Stress 2
SOLAR ANALYSIS ON GROUND
29
The Solar Analysis on the ground plane show the shadow effects generated by the surfaces. In Figure 43, the density of the red colored reflections is very low compared to Figures 44 & 45. This is due to the different heights of the models.
In Figure 44, around the surface, we can see the solid red parts which illustrate the solidity of the surface itself, where in Figure 43 and 45 we cannot see that much.
Fig.43 : Analysis 1/ Solar Analysis on Ground
The most dense surface is in Figure 45, resulting on the densest shadows being projected onto the ground.
Fig.44 : Analysis 2/ Solar Analysis on Ground
Weather File Source: London Wea Ctr St James Park - GBR Month: May Hours: 8:00 am /18:00 pm
Fig.45 : Analysis 3/ Solar Analysis on Ground
30
WIND PRESSURE ANALYSIS
Fig.46 : Analysis 1/ Wind Pressure (Pa) Plot
Fig.47 : Analysis 2/ Wind Pressure (Pa) Plot
Fig.48 : Analysis 3/ Wind Pressure (Pa) Plot Figures 46, 47 & 48 show the wind pressures on the surfaces in line plot graphs. From the resilliance of the graphs, we can see that the most regular one is the Analysis 2 (depends on the symmetrical geometry) and the most irregular one is the Analysis 1 as it has an asymmetrical geometric composition.
WIND VELOCITY ANALYSIS
Fig.49 : Analysis 1/ Wind Velocity (m/s) Diagram
Fig.50 : Analysis 2/ Wind Velocity (m/s) Diagram
Fig.51 : Analysis 3/ Wind Velocity (m/s) Diagram Figures 49, 50 & 51 show the wind velocities on the surfaces with the velocity vector representations (strips, particles and tubes respectively). From the movement of the velocity vectors on the graphs, we can see that the most trapped vectors are inside the Analysis 1, as the surface has the density on horizontal direction. The Analysis 2 is less dense compared to the others, where the Analysis 3 is the highest one and distributes the velocity vector in vertical directions.
31
32
STAGE 3 RESULTS
Fig.52 : Stage 3 Analyses Results of Solar/ Wind
33
34
STAGE 3 RESULTS Through the Stage 3, wind, solar and structural analysis were held in order to compare the surfaces’ performances. The maximum wind velocity on the surfaces for the Analysis 1, 2 & 3 were 14.362 m/s, 13.273 m/s and 21.270 m/s and the maximum pressures values were Pa, 67.414 Pa, 76.676 Pa and 311.667 respectively. So the Analysis 3 showed the highest maximum pressure and velocity values on the surface. For the wind pressure line plot graphs, the most regular one appeared in the Analysis 2 due to its symmetrical geometry, and the most irregular one was the Analysis 1. In the wind velocity analysis, we saw the most trapped vectors inside the Analysis 1, In order to compare the three models we ran a structural analysis on Karamba plugin for Grasshopper software, considering Concrete C12/C15 as the building material and testing them only for their self-weight load. From looking at the results generated by the software, we can see the model chosen from Stage 1 is the one with the biggest Displacement and Principal Stress values. This is mainly due to its lack of supports next to the ground plane. By comparing the second and third model, we can see the Displacement and Principal Stress values are bigger for the third one than for the second one. This is due to the difference in height between them, being the third one more than two times higher than the second one. Even so, for the second model we can see the displacement is very big in the middle compared to the perimeter, whereas the third model is more equally distributed all over its geometry. In the solar analysis of the coloured diagram of the surface, we saw the variation of solar radiation density at certain parts. Analysis 3 is the highest one among the 3 Analysis, and the solar radiation is getting less on the below parts of the surface. The Solar Analysis applied on the ground plane show the shadow casting effects of the surfaces. The smallest height amongst the models is the one for Analysis 1, this could be appreciated from the density of the red coloured reflections in Figures 43, 44 & 45. The most dense surface is the one in Analysis 3, as the solar radiation probed to be the most dense on the ground and as well as for most solid parts on the Analysis 2.
35
36
STAGE 4: FINAL STAGE In order to simplify the third model, we decided to conduct a series of transformations to its geometry. The aim was to create an abstraction of the model, trying not to lose the information generated by the agents. That’s why we chose to create a new continuous surface by joining the main clusters in the agents’ trails. This new surface was a continuous shell structure with the same proportions than the agent-based output, and a series of peaks standing out in the places where the highest intersection of elements occurred in the previous model. Even if this new model managed to keep most of the original characteristics present in the former model, it lost the permeability this one had. The new surface enclosed a space in a much more binary way, without letting any sunlight or wind go through it. That’s why we decided to keep the abstraction process, by subdividing the surface into a grid of U and V lines. With this new grid structuring the model, we piped the curves using a constant diameter of 60 mm. We felt the resulting model fulfilled all our expectations to keep the main features present in the original model as well as becoming simpler in terms of structure and material organization.
Fig.53 : Final Pavillion Geometry Rationalization Stage
Agents Behaviour Output
UV Subdivision
Average Surface Between Clusters
Final Geometry
Abstraction
37 This seminar course showed us a new way to carry out a design process. We were able to extract a social insect behaviour and turn it into a system capable of generating multiple outcomes. This process could eventually be repeated several times, working with different insect behaviours in order to develop new systems. Each one of the outcomes generated by our system has the potential to become a design model, but none of them is the final design. A huge amount of analysis and design development is needed to turn any of these starting points into a digital prototype capable of being built and inhabited. This makes this process no different than any other design process, where there is a huge amount of work between the starting point and the end of the construction phase. Even if we conducted a few environmental and structural tests, and our model is getting a little closer to becoming a digital prototype, much more design development is needed for that to happen. We would have to make material tests to check how this geometry would work for different materials, develop joints between the parts, run more structural analysis loading the structure with different loads coming from different positions according to the environmental forces acting on it, check the development and building cost of our project corresponds to the budget, etc. This agent-based system works as an initial form finding algorithm, but other variables have to be taken into account to turn this forms into a final model.
38
REFERENCES 1. Heylighen, Francis. «Stigmergy as a Universal Coordination Mechanism I: Definition and Components». Cognitive Systems Research 38 (June 2016): 4-13. https://www.sciencedirect.com/science/article/pii/ S1389041715000327?via%3Dihub 2. Dorigo, Marco; Bonabeau, Eric; Theraulaz, Guy. «Ant Algorithms and Stigmergy». Future Generation Computer Systems 16, n.o 8 (June 2000): 851-71. https://www.sciencedirect.com/science/article/pii/ S0167739X0000042X?via%3Dihub 3. Diez, Lise; Lejeune, Philippe; Detrain, Claire. «Keep the Nest Clean: Survival Advantages of Corpse Removal in Ants». Biology Letters 10, n.o 7 (July 31st 2014): 20140306. https://royalsocietypublishing.org/doi/10.1098/ rsbl.2014.0306 4. Martin, Marc; Chopard, Bastien ; Albuquerque, Paul. «Formation of an Ant Cemetery: Swarm Intelligence or Statistical Accident?» Future Generation Computer Systems 18, n.o 7 (august 2002): 951-59. https://doi. org/10.1016/S0167-739X(02)00074-2. 5. Choe, Dong-Hwan; Millar, Jocelyn G.; Rust, Michael K. «Chemical signals associated with life inhibit necrophoresis in Argentine ants». PNAS 106 (20) (May 19, 2009): 8251-8255. https://doi.org/10.1073/pnas.0901270106 6. Reynolds, Craig W. «Steering Behaviors For Autonomous Characters». Sony Computer Entertainment America. https://www.red.com/cwr/