Unlearning Phenomena in Co-evolution of Non-uniform Cellular Automata

This paper presents results of a study of a genetic algorithm, designed to evolve cellular automata for solving a given problem. Evolution is performed between the cell’s update-rules in a local manner, allowing for easy parallelization. As a case study, the algorithm was applied to the density classification problem: classifying any given initial configuration according to the percentage of 1-valued cells. The main result presented in this paper is an ’unlearning’ phenomenon: highly fit solutions are generated by the algorithm, only to be ’unlearned’ and completely disappear as the evolutionary run continues.