Use of simulation-aided reinforcement learning for optimal scheduling of operations in industrial plants

In this paper, we present an algorithm based on reinforcement learning for scheduling the operations of an industrial plant which is modeled as a network of machines on a directed acyclic graph. The algorithm is assumed to have access to a high-fidelity simulator of the plant, but not a mathematical model. The algorithm is designed to optimize an objective function over a moving window, similar to receding horizon control, for a typical industrial plant which converts raw material into finished products. The delivery schedule for the incoming raw material is assumed to be known but subject to uncertainty. A novel feature of our technique is the use of schedule moments to train the algorithm to handle a large class of incoming delivery schedules.