A Flexible Reinforcement Learning Framework to Implement Cradle-to-Cradle in Early Design Stages

Reinforcement Learning (RL) is a paradigm in Machine Learning (ML), along with Supervised Learning and Unsupervised Learning, that aims to create Artificial Intelligence (AI) agents that can take decisions in complex and uncertain environments, with the goal of maximizing their long-term benefit. Although it has not gained as much research interest in the AEC industry in recent years as other ML and optimization techniques, RL has been responsible for recent major scientific breakthroughs, such as Deep Mind's AlphaGo and AlphaFold algorithms. However, due the singularity of the problems and challenges of the AEC industry in contrast to the reduced number of benchmark environments and games in which new RL algorithms are commonly tested, little progress has been noticed so far towards the implementation of RL in this sector. This paper presents the development of the new Grasshopper plugin "Pug" to implement RL in Grasshopper in order to serve as a flexible framework to efficiently tackle diverse optimization problems in architecture with special focus on cradle-to-cradle problems based on material circularity. The components of the plugin are introduced, the workflows and principles to train AI agents in Grasshopper are explained and components related to material circularity are presented too. This new plugin is used to solve two RL problems related to the circularity and reuse of steel, timber and bamboo elements. The results are discussed and compared to traditional computational approaches such as genetic algorithms and heuristic rules.