Solving batch process scheduling/planning tasks using reinforcement learning

The complex and dynamic nature of shop-floor environments, coupled with unpredictable market demands, makes batch plant's reactivity a crucial management issue. In this work, reinforcement learning and a repair-based search strategy are integrated together in a learning problem-solver for scheduling/planning tasks. The overall design of the learning algorithm is based on a state-space search perspective in which the associated optimization problem is solved by starting in some initial infeasible solution and then proceeding to progressively repair intermediate solutions until a feasible one is found. A key component to accelerate search in the state space is learning an evaluation function that accumulates context-dependent knowledge about the goodness of applying a small set of repair operators so that future (re)scheduling and (re)planning problems can be solved with less effort. A demonstrative example is used to illustrate the importance of improving plant reactivity by learning to repair infeasible plans and schedules. (C) 1999 Elsevier Science Ltd.