Collaborative housing agents (MSc thesis)

Project developed ofr the thesis and final dissertation for the MSc Architectural Computation at The Bartlett - UCL

Course 2020/21

Tutors: Tommaso Casucci, Sherif Tarabishy

Developed in Unity using Unity ML-Agents release 18

ABSTRACT

Multi-Agent Reinforcement Learning (MARL) has enjoyed great success in decision-making domains such as robotics, games, or autonomous driving. Simultaneously, housing prices have continued to rise, contributing to social inequality. This thesis seeks to study the use of Deep Reinforcement Learning (DRL) methods for designing collaborative housing aggregations. It examines whether intelligent agents can cooperate and learn behaviours that create complex systems. The research explores the training and evaluation phases, using game engines to compute interactive environments. Distilling housing problems into sets of simple rules, it prioritises urban parameters such as height, sun-light, protected green areas or floor-to-area ratio (FAR). Computationally, the training phase comprises several configurations, based on the novel MultiAgent POsthumous Credit Assignment (MA-POCA) that uses decentralised agents and a central critic. The resulting behaviours are evaluated using the same conditions and comparing to understanding the rules that yield more efficient housing aggregations. After training, key concerns are maximising occupancy ratios, FAR and time to compute aggregations. Tools that enable decentralised mass housing facilitate access to high-quality, affordable homes for those excluded from homeownership. This study suggests an innovative use of MARL systems for collaborative housing where agents continuously interact to meet their briefs. The proposed computation allows for the implementation of additional rules and can be applied to other architectural problems. Keywords: deep reinforcement learning, multi-agent, MARL, DRL, MA-POCA, collaborative, housing, decentralised, agents, central-critic, housing