Prior-informed adaptive multi-objective graph reinforcement learning for lysine fed-batch fermentation process

Abstract

The optimization and control of modern industrial processes are increasingly complex, often necessitating fine trade-offs among multiple conflicting objectives. Multi-Objective Reinforcement Learning (MORL) offers a highly promising paradigm for addressing such problems. However, existing MORL methods commonly face challenges in effectively integrating domain prior knowledge and exhibit low data efficiency. To tackle these challenges, a novel MORL framework named Pri-AG is proposed in this paper. Pri-AG utilizes Graph Convolutional Networks (GCNs) to effectively integrate an initial graph topology derived from domain knowledge. Furthermore, a hypernetwork is employed to dynamically adjust graph edge weights according to preferences, thereby enhancing learning efficiency and improving model interpretability. The proposed Pri-AG framework was systematically validated on a typical lysine fed-batch fermentation process simulation. Extensive experimental results demonstrate that Pri-AG enhances sample utilization efficiency, accelerates policy convergence, and surpasses other MORL benchmark algorithms across most performance metrics.