Application of Reinforcement Learning to the Westenberger–Kallrath Problem

The nature of chemical processes imposes multiple challenges on production planning and scheduling. The Westenberger–Kallrath (WK) problem which was published in 2002 still serves as a benchmark for that industry. In the past, mathematical models and solution approaches made use of mixed-integer (linear) programming (MI(L)P) methods or metaheuristics. Nowadays, algorithmic advances in artificial intelligence (AI) provide new opportunities for integrated modeling and solution methods. In this research, we investigate the application of reinforcement learning (RL) and propose a novel approach to solve the WK problem. Specifically, we develop a material requirements planning (MRP) and batching heuristic to preprocess the problem data and create chains of production orders that can be scheduled independently. Next, we apply RL algorithms to train an agent to schedule the chains following the objective of minimizing the makespan of the complete schedule. We detect modeling and implementation challenges arising from first experiments.