COGNITIVE ASSEMBLAGES Sharpening machine sensibility for architectural design of spatial assemblages through Wave Function Collapse and Reinforcement Learning
Alessandro Mintrone University of Bologna
Introduction Theory overview
Is it possible connect, into the distributed lattice of decisions of the design process, sufficiently complex non-human cognitions capable of claiming part of the authorship, generating spaces, places and everything comprised in the larger spectrum of architecture?
In this page. Manus, robotic installation by Medeline Gannon, at the World Economic Forum.
Cognitive Assemblages
Artificial Neural Networks
Artificial Neural Networks Artificial Intelligence
In this page. AI classifiction scheme, according to Ian Goodfellow in his book Deep Learning Cognitive Assemblages
Artificial Neural Networks Structure
Artificial Neural Networks learn how to map an array of inputs into an array of outputs, passing thru a defined number of internal layers, each one composed by a defined number of units connected to the adjacent layers. What is learned during the training process are the weights of these connections. In this page. Neural network model with a single hidden layer Cognitive Assemblages
Artificial Neural Networks Example
In this page. Example of a classifier based on a Convolutional Neural Network Cognitive Assemblages
Artificial Neural Networks Latent Space
In this page. Representation of the internal latent space, visualizing categorial boundaries Cognitive Assemblages
Artificial Neural Networks Tools
High level features A neural network can deal with high level features, such as the overral shape of the head of the skin’s texture, capturing non-linear relationships between these features.
In this page. Two images generated within the project: This person does not exist. Cognitive Assemblages
Wave Function Collapse
Wave Function Collapse Overview
Entropy Analysys
Collapsing Point Selection
Possible Tiles [0, 1, 1, 0, ..., 1, 0]
Contraddiction
Tiles Selection
Diffusion
Algorithmic process
Tileset and resulting assemblage
Cognitive Assemblages
Wave Function Collapse Connection rules
Possible Tiles [0, 1, 1, 0, ..., 1, 0]
Contraddiction
Tiles Selection
Single tiles
Permutations
Contraddictions
Assemblage
Possible connections
Forbidden connections
Cognitive Assemblages
Wave Function Collapse Information diffusion
Information diffusion Tiles Selection
Diffusion
Set di tiles
Cognitive Assemblages
Tile Set 02 Tridimensional oriented
Cognitive Assemblages
Tile Set 02 Tridimensional oriented
Cognitive Assemblages
Tile Set 03 Tridimensional cubic
Cognitive Assemblages
Tile Set 03 Tridimensional cubic
Cognitive Assemblages
Tile Set 05 Tridimensional planar
Cognitive Assemblages
Tile Set 05 Tridimensional planar
Cognitive Assemblages
Reinforcement Learning
Reinforcement Learning Cognition and environment
With Reinforcement Learning the neural network learns from experience. Constructing a feedback-loop by acquiring informatins from the environment, selecting an action and, in this way, transforming the environment itself.
In this page. Feedback loop diagram of the agent interacting with the environment Cognitive Assemblages
Reinforcement Learning Cognition and environment
Axime ni is veniendi omnis etur solorerat ius et re voluptatquas ditene qui omnis aborit, quasperibus et haruptaqui ullorita cus nimint laceptaspid eum doluptat quidunto inum reptis entioribusam ullenditaqui aut providentur? At exceperum nulparu mquatiis moluptaesto omnimin ctionsectio. Bus dolorum explis aut pos unto quam e
In this page. AlphaGo, is the first computer program to defeat a professional Go player. The system combines Reinforcement Learning with a Montecarlo tree-search algorithm Cognitive Assemblages
Wave Function Collapse Intelligence
Entropy Analysys
Entropy Analysys
Collapsing Point Selection
Collapsing Point Selection
Possible Tiles
Possible Tiles
[0, 1, 1, 0, ..., 1, 0]
[0, 1, 1, 0, ..., 1, 0]
Contraddiction
Contraddiction
Tiles Selection
Tiles Selection
Diffusion
Neural Network
Diffusion
Algorithmic process
Neural Network inception
Cognitive Assemblages
Density
Gradient
Spatial Analysis
Spatial Analysis
1. Density 1 - |desiredDensity - currentDensity|
2. Gradient 1 - fieldValue
Cognitive Assemblages
Orientation
Structural Connectivity
Spatial Analysis
Spatial Analysis
3. Orientation allignedTiles / totalTiles
4. Structural Connectivity 1 - tilesUnderThreshold / totalTiles
Cognitive Assemblages
Spatial Connectivity
Planar Connectivity
Spatial Analysis
Spatial Analysis
5. Spatial Connectivity 1 - tilesUnderThreshold / totalTiles
6. Planar Connectivity 1 - tilesUnderThreshold / totalTiles
Cognitive Assemblages
Machine Learning Rete neurale
3 hidden layers
Input
output
0, 0, 1, . . . 0, 1
0, 0, 1, . . . 0, 1
256 units
256 units
256 units
Neural Network Model
Cognitive Assemblages
Machine Learning Analisi e visualizzazione
Parameters visualization on Tensorboard
Cognitive Assemblages
Training Heuristic comparison
The Neural Network obtains not only better results in terms of reaching a target density, but also determines non-linear behaviors that led to pattern generation.
Porbabilistic Heuristic Required density: 50% Actual density: 76%
Trained Neural Network Required density: 50% Actual density: 54%
Cognitive Assemblages
Training Density
Target value
Analysis Parameters
Cognitive Assemblages
Training Orientation
Target value
Analysis Parameters
Cognitive Assemblages
Training Connectivity
Target value
Analysis Parameters
Cognitive Assemblages
Training Density + Connectivity
Target value
Analysis Parameters
Cognitive Assemblages
Training Density + Orientation
Target value
Analysis Parameters
Cognitive Assemblages
Training Orientation + Connectivity
Target value
Analysis Parameters
Cognitive Assemblages
Training Global
Cognitive Assemblages
Discrete Assemblages
Assemblages Fabrication and automation
In this page. Designed by Moshe Safdiefor the 1967 Expo in MontrĂŠal, the Habitat use prefabricated parts and a combinatorial logic for spatial generation
Cognitive Assemblages
Assemblages Fabrication and automation
Ăˆ possibile inserire, nella rete di decisioni del processo progettuale, cognizioni non-umane sufficientemente complesse da assumere in parte la responsabilitĂ della creazione, per generare spazi, luoghi e ogni altro esito che è possibile chiamare architettura?
In questa pagina Installazione di Gilles Retsin per la biennale di architettura di Tallin (TAB) del 2017. Cognitive Assemblages
Assemblages Parts and connections
In this page. Case study, example of design and connection rules of the selected tileset
Tileset
Connection rules: corners
Connection rules: horizzontal parts
Connection rules: diagonal parts
Connection rules: vertical parts Cognitive Assemblages
Assemblages Comparative Analysis
Cognitive Assemblages
Assemblages Comparative Analysis
Cognitive Assemblages
Assemblages Comparative Analysis
Cognitive Assemblages
Assemblages Analysis overview
Voids analysis
Overview and analysis parameters
Connectivity Analysis
Cognitive Assemblages
Assemblages Analysis overview
Plan view
Autonomously generated internal views
Section view
Cognitive Assemblages
Assemblages Analysis
Cognitive Assemblages
Voids
Connectivity
Plan
Section
Assemblages Analysis
Cognitive Assemblages
Voids
Connectivity
Plan
Section
Assemblages Analysis
Cognitive Assemblages
Voids
Connectivity
Plan
Section
Assemblages Analysis
Cognitive Assemblages
Voids
Connectivity
Plan
Section
Case Study
Cognitive Assemblages
Case Study
Cognitive Assemblages
Case Study
Cognitive Assemblages
Cognitive Assemblages
Cognitive Assemblages
Cognitive Assemblages
Cognitive Assemblages
Cognitive Assemblages
Case Study
Cognitive Assemblages
Case Study
Cognitive Assemblages
Case Study
Cognitive Assemblages
Case Study
Cognitive Assemblages
Case Study
Cognitive Assemblages
COGNITIVE ASSEMBLAGES Alessandro Mintrone University of Bologna alessandro.mintrone@hotmail.com