Reinforcement Learning for Engineering Design Optimization: A Task-Oriented Critical Review

Reinforcement learning (RL) is increasingly being explored for engineering design optimization because of its ability to learn adaptive strategies through interaction and feedback. However, its wider application is still limited by high computational costs, strict engineering constraints, and difficulties in transferring learned strategies to new tasks. This review uses a task-oriented framework to connect engineering design tasks with Markov decision process formulations. A total of 197 references were retained for analysis, including direct engineering design studies, methodological studies, reviews, and contextual sources used to define the task-to-MDP framework. Commonly used RL algorithms, state and action representations, reward formulations, and constraint-handling methods are summarized across parametric, shape, and topology or graph-structured design. Applications involving multiple objectives, coupled disciplines, uncertainty, and joint optimization of design and control are also discussed. Finally, the suitability of RL for different design tasks and the remaining challenges are analyzed, and suggestions for future research are provided.