A Population Based Rewarding for Reinforcement Learning to Control Genetic Algorithms

The effectiveness of Genetic Algorithms (GA) heavily depends on the appropriate setting of its parameters. Moreover, optimal values for these parameters depend on both the type of GA and the application problem pattern and must be developed for each particular setting one by one. Therefore it requires special expertise and many experiments to validate the parameter setting. In order to solve this problem, a new method called "adaptive parameter control" was proposed, which adaptively controls parameters of an evolutionary algorithm. However, since this method just increases the selection probability of a search operator that generated a well evaluated individual, this is apt to be a shortsighted optimization method. On the contrary, a method is proposed to realize longsighted optimal parameter control of GA using Reinforcement Learning (RL). However, this method does neither consider the calculation cost of search operators nor population search characteristics of GA. Here, we propose a refined RL method for parameter control, in which (1) the reward decision rules are elaborately incorporated under the consideration of GA\´s population search characteristics and (2) the calculation cost of the search operator is taken into account. It is expected that this method can efficiently learn parameters to optimally select search operators of GA for approximately solving Traveling Salesman Problems (TSPs).