深層強化学習による連続変数と離散変数を有する補剛板の構造最適化

抄録

Optimizing stiffened panel structures used in ships and aircrafts are becoming increasingly important for reducing material costs while maintaining or enhancing structural strength. These structures require simultaneous optimization of continuous design variables (panel thicknesses) and discrete design variables (stiffener cross-sectional shapes), making gradient-based methods less applicable. Consequently, heuristic approaches such as Genetic Algorithms (GA) are often employed; however, GA typically imposes a high computational burden in large-scale optimization problems. To address this, the use of deep reinforcement learning (DRL) has been investigated for large-scale combinatorial optimization. Among DRL approaches, Double Deep Q-Network (DDQN) has been reported as effective, yet its application to structural optimization involving high computational demands from Finite Element Analysis (FEA) remains limited. To overcome such an environment, in this study, an optimization flow for structural optimization is considered, and states, actions, and rewards appropriately representing design variables, constraint conditions, and objective functions are discussed. In addition, this study also proposes incorporating an elite-preservation algorithm into DDQN to reduce the computational load of structural optimization. Experimental results show that the proposed method yields designs under various load conditions that are up to 6% lighter than those obtained using GA, with a computational time reduction of approximately 81%. These findings confirm the feasibility of efficient and effective optimization for stiffened panel structures and suggest potential benefits in cost reduction and performance enhancement in future structural designs.