A Graph-Based Grammar for Structural Design using Deep Reinforcement Learning

This paper introduces a graph-based grammar method that combines the Combinatorial Equilibrium Modeling (CEM) form-finding approach and Deep Reinforcement Learning (RL). We formalize the design process of the CEM as a Markov Decision Process (MDP) that can act as an environment for an RL agent. This discrete step-wise design process allows both geometrical and topological manipulation of structural designs through a structural grammar consisting of a specific set of design actions. All design actions achieve design state transitions that preserve not only equilibrium but also the validity of the input structural topology. This guarantees that all design states are fully interactable parametric models, enabling intuitive manipulation of intermediate structural solutions by human designers. Parameters are represented by the input features of the CEM that relate to physical quantities of the structure such as desired edge/member lengths and forces. The combination of the top-down constraints of the CEM and the bottom-up design actions of the MDP results in a cross-typological design space that can be collaboratively explored by a human designer and an RL agent.     «

This paper introduces a graph-based grammar method that combines the Combinatorial Equilibrium Modeling (CEM) form-finding approach and Deep Reinforcement Learning (RL). We formalize the design process of the CEM as a Markov Decision Process (MDP) that can act as an environment for an RL agent. This discrete step-wise design process allows both geometrical and topological manipulation of structural designs through a structural grammar consisting of a specific set of design actions. All design ac...     »