Enhancing architectural space layout design by pretraining deep reinforcement learning agents

Abstract

Space layout design is a fundamental yet complex problem in architecture. This task’s inherent complexity, arising from the need to balance numerous geometric configurations and topological relations while adhering to specific constraints, poses significant challenges. Recent advancements in deep reinforcement learning have shown promise in addressing similar planning problems, suggesting its potential utility for innovative space layout solutions. However, a critical limitation of deep reinforcement learning is its struggle with generalizing learned strategies to unseen scenarios. In the context of architectural design, this limitation could prevent deep reinforcement learning from being a scalable design method. Pretraining has emerged as a transformative strategy within the field of artificial intelligence, especially in the realm of foundational models, to enhance the generalization capabilities of learning algorithms. While pretraining is being the central focus of this paper, our approach diverges from conventional pretraining methods that focus on pixel-level design of layouts as is in diffusion based model. Instead, we propose an architectural simulation of space layout design that could embody the multifaceted essence of architectural design. To this end, we have developed a space layout simulator called SpaceLayoutGym that serves dual purposes: first, as an environment for the reinforcement learning agent to interact with and learn the intricacies of design, and second, as a tool for generating a dataset of design scenarios and their corresponding design solutions for model pretraining. We then used imitation learning to pretrain the agent on the generated training design scenarios. This process is being followed by a fine-tuning phase by using proximal policy optimization algorithm on new design scenarios. Our results demonstrate that pretraining can enhance the generalization capabilities of deep reinforcement learning in space layout design, paving the way for more adaptable and scalable artificial intelligence-aided architectural design.