An Academy of Spatial Agents

Motivated by situated cognition in architectural design, this research investigates a computational approach for spatial synthesis that supports designers inside the generative loop with fine-grained interaction. After reviewing and creating a taxonomy of classic paradigms of generative modeling, we present a research guideline for interactive spatial synthesis. This guideline comprehends control of problem parameters, real-time interaction, stepby-step and fine-granular construction of the spatial representations, event-driven computing, reduced assumption about control strategies, and simultaneously addressing multiple spatial goals. 

To test the feasibility of this research guideline, we investigate multi-agent spatial synthesis (MASS), which comprehends a collection of discrete and autonomous entities referred to as spatial agents that interact in a shared environment. By distributing the computation, MASS benefits from robustness and scalability and can enable designers to engage in interactive simulations to explore spatial patterns in large state spaces under uncertainty. Besides, similarly to agent-based models, MASS can potentially model the dynamics of complex systems and display qualities such as emergence and complexity. Nevertheless, there is no clear methodology to design agents that can manage conflicts with realistic spatial synthesis goals. Existing prototypes for MASS typically rely on trends such as bioinspiration and physics-simulation and often on hybridization with global solutions, such as metaheuristics. 

We address this bottleneck of agent design in a MASS research artifact by using multi-agent reinforcement learning as a training methodology for spatial agents. It enabled creating agent policies that address multiple architectural goals and that support fine-grained interaction with reduced dependency on initial design states and order of operations. The resulting artifact is based on training an ecology of spatial agents with parameter sharing and a parameterized reward function to support domain knowledge and enable heterogeneous agent behaviors. The MASS artifact was evaluated in multiple design cases: a house design in different site conditions, a museum on a large interstitial site, and a speculative design of a house complex on a large empty site. Overall, the agent policies satisfied custom spatial goals and produced consistent morphological patterns for the different design cases. 

After the evaluation of the MASS artifact, we integrated it to an interactive simulation workflow to support designers in the generative loop with access to fine-grained interaction, event-driven computing, and other properties described in the initial guideline. The success of the design artifact and the implementation of the interactive simulation workflow are proofs by construction of the feasibility of the proposed approach to design spatial agents for interactive spatial synthesis. Besides, they open doors to different lines of research, such as agentbased architectural morphology, spatial synthesis of complex spaces based on ecologies of agents, and mainly, the evaluation of situational and collaborative design behavior with architects.