Agent of deep reinforcement learning for multi-objective arch dam shape design

Abstract

Shape design is the dominant task of arch dam construction, involving significant computational costs. Conventional approaches are largely manual and experience driven. Though surrogate-assisted methods accelerate the procedure, the reusable “optimization policy” is ignored. Inspired by the cyclical interactions between designers and experts in real-world engineering, a deep reinforcement learning (DRL) framework is proposed for automated and intelligent arch dam shape optimization. The framework models the arch dam design as a DRL task and employs the Soft Actor–Critic algorithm to train the agent, with Gaussian process surrogate models accelerating the procedure. A weight-vector-based transfer learning strategy is introduced to generalize the framework to solve multi-objective problems. The framework is implemented on a real-world arch dam, and the results demonstrate that the agent effectively learns an optimization policy and generates a high-quality Pareto front. The selected optimal shape achieved 12.5% and 25.87% reductions in dam volume and tensile volume, respectively, demonstrating enhanced economic efficiency and structural safety. The same methodology can be widely applied to other engineering structure designs and has the potential to drive transformative advances in the engineering community.