Fitness Landscapes of Evolved Apoptotic Cellular Automata

Daniel Ashlock and Carolyn Pugh
Submitted to the 2013 Congress on Evolutionary Computation

Abstract PDF eprint

Cellular automata are known to be capable of Turing-complete computation and yet programming them to do particular tasks can be quite daunting. In this paper we use single parent crossover as a means of transferring information between successive evolving populations to create rules for cellular automata that have proscribed shapes. The proscription of regions where the automata are permitted to grow is the reason they are called bonsai automata. This work follows earlier work on apoptotic cellular automata that simply exhibit self-limited growth. The correct choice of single parents permits enormous improvement in the performance of evolutionary algorithms searching for automata that satisfy particular bonsai templates. In this study, we demonstrate that single parent techniques make meeting shape constraints on the growth of CAs possible at all in some cases. This study also introduces range niche specialization to control problems with the cloning of ancestors used for single parent crossover in an evolving population. This study demonstrates that different bonsai shapes have highly variable difficulty. It is also shown that automata evolved to satisfy one bonsai template may be needed to enable, via single parent crossover, solutions for another template. The use of bonsai techniques yields many automata not found during studies of apoptotic automata demonstrating that the technique encourages exploration of different parts of the fitness landscape.