Adaptive Utility Multi-Objective Reinforcement Learning for Optimizing Batch-Fed Fermentation Processes

In industrial applications, many scenarios involve multiple conflicting objectives, such as yield, raw material consumption, and environmental protection in chemical production. To address this challenge, Multi-Objective Reinforcement Learning (MORL) provides an effective solution by learning Pareto-optimal policies through multi-objective vector rewards. Despite these advances, a gap remains in accurately capturing and incorporating user preferences into the MORL framework. In response to this, a novel Adaptive Utility MORL algorithm (AU-MORL) is proposed to improve policy optimization in continuous fed-batch fermentation processes. This algorithm adaptively adjusts the utility between multiple objectives and uses reinforcement learning to generate policies aligned with user preferences. Extensive validation in the fermentation simulation environment demonstrates that the algorithm achieves a denser pareto front that more closely approximates the true front, while the resulting policies effectively align with user preferences.