Studio: FLOW 2019/2020
Project: Evolving Architecture – Adaptation of Artificial Neural Networks in Complex Dynamic Environments
Author: Karolína Kotnour, Ing. arch.
Supervisor: Miloš Florian, doc. Ing. arch. PhD.
Cooperation: Institute for Research in Science and Art, Robert B. Lisek, MFA M.Phil. Ph.D.
Supported from Student Grant Competition: SGS19/117/OHK1/2T/15
Meta-Evolver is a tool for a visual representation of the various models and generated resulting correlated layers. The 3D environment for testing dynamic spatial adaptation, where the environment is composed of algorithms and parametric definition.
The proposed research uses advanced AI methods (meta-learning) and cutting-edge technologies including immersive environments and virtual reality (VR) to offer innovative methods of architectural creation. The ability to continuously learn and adapt from limited experience in a dynamic environment is an important milestone on the path towards building interactive spaces in modern architecture. Besides randomly generated data of the dynamic environment, the dynamic environment was defined as a sound space and the neurofeedback as a training data layer for the analysis of the complex environment and agent continuous adaptation. We developed the tool Meta-Evolver for testing spatial adaptation in dynamic environments and integrated the ability for interaction with a human user.
Method of the research
The framework for the adaptive agent-based model for dynamic environments based on data of generated random numbers and soundscape was defined. We outlined and established the architectural strategy of the multiplatform system for generative modeling based on various input datasets. Prepared the framework for a visual representation of the various models and generated resulting correlated layers.
Development of Meta-Evolver
The main task was adapting an agent to new environments and create a new multi-agent environment as well as architecture for testing various aspects of continuous adaptation. The whole model was parameterized and the communication protocols were integrated. The method was to present dynamics as a sequence of tasks and train agents to use the dependencies between successive tasks. We created a meta-learning model for the problem of continuous adaptation of an artificial actor/agent in a complex dynamic environment. The observation-based research on these generated correlations was conducted and defined the possible dispositions of forming patterns and structures, the model can be applied to various environments and after pre-training of agents can effectively adapt and, generate architectural dispositions, structures, and environments.
The parallel task of my research was the problem of creating virtual interactive environments (VE). I was focused on the roles of Presence, Flow, Immersion, and Interactivity. Presence is defined as the subjective experience of being in one place or environment, even when one is physically situated in another. Presence is a normal awareness phenomenon that requires directed attention and is based on the interaction between sensory stimulation, environmental factors that encourage involvement and enable immersion. Flow is a state of experience where someone is completely absorbed and immersed in an activity. I researched relations between presence, flow, immersion, and interactivity, e.g. how interactivity and sound spatialization improves the experience of presence.
The three different and complementary 3D environments and experiments: 1) adaptation in a dynamic environment created by changes in the structure of the parametrized environment; 2) adaptation in a multi-agent environment created by the presence of multiple learning actors (interdepended datasets, transformation matrices), and 3) adaptation in a dynamic environment created by the interaction of a human user with an adaptive artificial agent. The immersive dynamic environment is created by using virtual reality (VR) and sound synthesis. The model keeps the transformation of 3D objects and sound synthesis as synchronous processes.
The Meta-Evolver and ML algorithms and parametrized environment were made accessible at
https://github.com/kaiakk/Meta-Evolver
The documentation of the 3D dynamic environment and machine learning algorithms is accessible at