Acquisition of AGV Control Rules Using Profit Sharing Method and Evaluation of the Rules

Abstract

When a simulation-based approach is adopted in order to design production systems under uncertain conditions, a design support environment that generates simulators of various configurations automatically is indispensable. Also, such environment is required to devise suitable operating rules of each design candidate at the same time.

In the previous work, AGVs control rules were obtained by means of reinforcement learning method. Concretely, Profit sharing was adopted and the computer-based support system was developed by Java language by means of Object-Oriented technique. In this work, the state space of AGVs were extended so as to apply this approach to an AGVS with three or more AGVs. Also, an attempt was made to apply the obtained rule set to an AGVS with four or more AGVs. A series of experiments was carried out using this system in order to confirm whether proper rules for various path topologies and various number of AGVs can be learned, and evaluate the effectiveness of the rules.