Efficient Design-and-Control Automation with Reinforcement Learning and Adaptive Exploration

EDiSon (Efficient Design and Stable Control) casts design optimization as a multi-step MDP and learns design and control jointly: a design policy proposes a structure step by step while a control policy operates it, both trained with deep RL against a reward that scores design quality. A design memory drives adaptive exploration — the agent regulates between building a design from scratch and replaying a stored high-quality design to refine it, which balances exploration against exploitation and stabilises control-policy learning. Because the formulation is domain-agnostic, the same method covers robot morphology design and Tetris-based design, and targets automated materials discovery.