Application of Reinforcement Learning for the Generation of an Assembly Plant Entry Control Policy

The generation of an entry control policy for an assembly plant using a reinforcement learning agent is investigated. The assembly plan studied consists of ten workstations and produces three types of products. The objective of the entry control policy is to produce a given production mix within a planning horizon, while following a given production mix. Due to the large state space, a function approximator, based on a neural network, is used to model the long-term reward function. The schedules generated by the trained agent are compared to those produced by a deterministic heuristic control policy that has been developed for this assembly plant. Simulation results show that the reinforcement learning agent produces production plans that achieve better productivity than the heuristic controller under tight planning horizons, generating sub-optimal yet acceptable production mix balance.