Selective Generalization of CMAC for Q-Learning and its Application to Layout Planning of Chemical Plants

This paper proposes a modified design method of CMAC integrated in a reinforcement learning system to solve the allocation problem for the chemical plant. In the proposed method, the generalization of CMAC is selectively settled for each measure according to characteristics of measures. This feature is efficacious to improve the learning performance of the system. In application examples, by using the proposed method, the elements of the chemical plant are placed by choosing the position having the best evaluation that is obtained by adequate learning iteration. Then this procedure gives an allocation plan with minimized risk and maximized efficiency. In addition, rotated and/or shifted plants that have the same layout can be identically recognized, so that the learning performance can be improved.