Multi-objective optimal control of lysine fed-batch based on reinforcement learning

Reinforcement learning has been increasingly used in solving optimal control problems in recent years. The lysine fed-batch fermentation represents a classic problem in optimal control, and reinforcement learning has become a new direction for its solution. To maximize the yield and productivity of lysine, a constrained two-objective optimal control model is proposed. For this two-objective optimal control problem, first, it is transformed into an optimal control problem within a fixed time domain through a time-domain transformation; then, the penalty function method is employed to handle the constraints; Finally, an experimental environment for the fermentation process is designed in the OpenAI Gym library. The agent is trained extensively using the PPO algorithm to enable it to effectively adjust the feeding strategy during the fermentation process to optimize the target production index. The experimental results show that reinforcement learning significantly improves the efficiency of fermentation.