Improved evolutionary design for rule-changing cellular automata based on the difficulty of problems

This paper describes a method to promote the evolution of the transition rules of cellular automata using a genetic algorithm. We previously proposed the evolutionary design of a cellular automaton in which an applied rule changes with time. This method encodes a rule and the number of times the rule is applied as a chromosome. In this paper, we describe the improvement of the method and analyze rules obtained using the Lambda parameter defined by Langton. The difficulty of test problems in an evolutionary process is adjusted so as to obtain a rule which performs the density classification task with high probability. Experiments using ten-thousand randomly generated tasks have shown that the proposed method performs better than the previous method.