Evolution of Rules for Non-Uniform Cellular Automata using a Simple Genetic Algorithm

Diversity of computational rules would appear to be beneficial. After all being able to apply a combination of different rules to a particular problem has at least as much power as applying a single rule. However, it takes more time and effort to find multiple rules that work well together than it does to find one rule. We investigated a particular instantiation of this trade-off by using a simple genetic algorithm (GA) to evolve rules for a non-uniform cellular automata (CA) which is being used to solve the density classification problem. Our hypothesis was that if we controlled the number of rules the GA was evolving, the GA would have the best performance given computational effort for a medium number of rules, i.e. that two rules might surpass both one rule and three rules. Our initial results indicate that having two rules often does as well as, and occasionally outperforms a single rule. Moreover, three or more rules always does worse than two rules given the amount of computational effort that we gave to the GA. However, all of these results are dependent on the size of the search space. Track: Genetic Algorithms