A reinforcement learning approach to support setup decisions in distributed manufacturing systems

A reinforcement learning approach to specifying payoffs for setup games is presented. Setup games are normal form, noncooperative games used by heterarchical machine controllers to evaluate reconfiguration decisions. While past work utilizing heuristic measures to approximate the effect of setup decisions has demonstrated promising performance, the lack of an accurate long-term model of system dynamics in these heuristic approaches limits their usefulness. The reinforcement learning approach iteratively learns the long term costs of setup decisions, accounting for both immediate decision effects and the effects of likely downstream decisions.