Feature-Conditioned Reinforcement Learning for Generalizable Engineering Optimization: Benchmarking on Multimodal Test Functions

Generalization is essential for accelerating optimization workflow, enhancing scalability and reducing computational cost. This paper presents a novel Feature-Conditioned Reinforcement Learning (FC-RL) method for engineering optimization, which enables generalization across diverse problem scenarios by conditioning the RL policy on explicit problem features. Unlike traditional optimization methods that require hyperparameter tuning for each new task, FC-RL leverages a meta-network to condition its behavior dynamically through Feature-wise Linear Modulation (FiLM). The experiments conducted across six multi-modal benchmark functions with varying dimensions, search-space, and global optimum shifts demonstrate that FC-RL often achieves global convergence without the need of retraining or retuning, while producing consistent and reproducible results due to its deterministic policy. Although FC-RL’s performance depends on sufficient training diversity and is limited to novel scenarios within the known environments, these findings indicate that the proposed method is particularly well-suited for repetitive engineering optimization tasks involving variable problem configurations across multiple systems.