Abstract
Batch production scheduling of resources to meet fluctuating product demand is a critical topic in the process industry. Existing optimisation approaches, based on heuristic and exact methods, trade off solution optimality and scalability to large problems. In this work, we investigate deep reinforcement learning as a powerful alternative in order to learn heuristics for batch scheduling. We formulate the batch scheduling problem as a Markov decision process operating on a state-task network representation encoded using graph neural networks, capturing relevant structural inductive biases. We propose a centralised training with decentralised execution architecture, in which agents placed on machines individually choose which tasks to complete using a global view of the network, cooperating towards task schedules that optimise the final production quantity. Preliminary results demonstrate that the proposed end-to-end framework learns to construct task schedules comparable to the optimal solution on small instances unseen during training, exhibiting strong potential for extension to more general graph structures and better scalability.