Artificial intelligence (AI) algorithms are gaining increasing popularity in the domain of architecture, urban design, and landscape architecture. However, recent advancements using image-based approaches including generative adversarial neural networks do not incorporate human-centric evaluation metrics and are prone to potential bias embedded in the dataset researchers used to train AI agents. Moreover, the outcomes of such approaches are pixelated images that are not directly useable as architectural designs.
Inspired by enactive learning in developmental psychology, the machine learning community has developed increasingly powerful AI agents that learn emergent behavior through unsupervised approaches such as self-play or actor-critic that do not rely on human heuristic datasets.
Therefore, I propose Enactive Genesis, a novel environment to train generative architecture through reinforcement learning that uses human-centric evaluation metrics. The environment is composed of three parts, each part being an autonomous agent that achieves certain tasks in the reinforcement learning process:
Assemble Bot: An open-source developer pack that researchers could use to create generic architectural objects using a universal BIM grammar.
Navigation Bot: An open-source developer pack to generate autonomous quasi-human navigation in the space generated by component 1. The generated data will be used as reward metrics for optimize generation process.
Critic Bot: An open-source developer pack to train intelligent agents to generate and modify grammar in component 1 to optimize reward function according to human-centric evaluation metrics in step 2.
