Scheduling individualized products with reinforcement learning to reduce worker stress

In recent years Reinforcement Learning (RL) has been used successfully in the domain of production planning and control. RL agents provide a powerful alternative to dispatching rules as they are capable of learning complex policies and can react to dynamic environments. This work tackles a real-world use case similar to a flow shop scheduling problem with multiple objectives. While previous studies often considered economic objectives, this work added the consideration of reducing the worker stress. An RL agent is trained to solve the problem in almost real time. The agent is trained using genetic algorithm solutions as reference to overcome limitations to multi-objective agents in existing works. Furthermore, clustering algorithms are utilized for data preprocessing. This allows the use of simple linear networks instead of more complex Transformer or LSTM Networks. The method is tested on a set of real problem instances, testing the performance of the new reward method and different number of clusters. The performance for both objectives is similar to a genetic algorithm implementation, while drastically reducing the computation time.