12:00 Wed 16 Oct 1996, WeH 7220 IMPROVING THE QUALITY OF INDUSTRIAL SIMULATION USING REINFORCEMENT LEARNING ----------------------------------------------- Sridhar Mahadevan Department of Computer Science and Engineering University of South Florida Simulation is an extremely important tool for designing and analyzing complex industrial systems. It has become widely accepted as a low-cost method for testing systems and policies before incurring actual implementation costs. However, existing simulation software is fundamentally limited to evaluating fixed user-provided policies. This talk will describe an ongoing project on combining reinforcement learning with discrete-event simulation to provide a better tool for industrial systems design. A new model-free reinforcement learning algorithm for semi-Markov decision processes (called SMART) will be illustrated, using two commercial simulation packages (ARENA and CSIM). Empirical results will be presented for several versions of a machine maintenance problem. (This is joint work with Tapas Das, Abhijit Gosavi, and Nicholas Marchalleck)