Abstract
Progressive evolution is one of the promising methods to improve convergency of genetic algorithms (GAs). It can find the solution with fewer iterations than conventional GAs by giving step-by-step subgoals that can be easily reached, not by hard task of searching directly for a solution in a large space. Conventional progressive evolution methods, however, have a serious problem that each subgoal must be set manually. Furthermore, they require extensive knowledge about a given problem such as the landscape of the search space, because of the obligatory manually set subgoals. For these problems, we previously proposed a progressive evolution method that can set subgoals autonomously. Yet, it still requires knowledge about problems such as the problem solving level. In this paper, we propose a new progressive evolution method that can set subgoals autonomously with little knowledge of the problem. In our method, we define the next subgoal as the common properties in an elite population of the current generation. Setting subgoals only requires the properties of individual phenotypes. Thus, little knowledge about the problem is required. We applied our method to the 6-multiplexor problem and the action control circuit for artificial ants to evaluate its searching ability. The experimental result confirms that our method performs progressive evolution with good convergency by appropriately setting subgoals.
