Evolution is commonly viewed as a process of hill climbing on a fitness landscape. A major problem with such a view is the presence of local optima; sub-optimal regions of the landscape from which no further progress is possible. There is an increasing amount of evidence, however, that the presence of large degrees of redundancy in the genome may alleviate this problem through the creation of neutral networks; sets of genotypes at the same level of fitness that are connected by single point mutations. These networks allow drift at the same fitness level and hence may increase the reliability of the evolutionary process by allowing the exploration of larger portions of genotype space. The presence, or otherwise, of genetic redundancy couldthus be an important concern in the design of artificial evolutionary systems. This paper explores the effects of genetic redundancy in the context of an evolutionary robotics experiment. Neural network control systems are evolved for a simple navigation task and the speed and reliability of the evolutionary process ascertained for differing levels of redundancy. Evolutionary progress is found to halt far more readily as the degree of redundancy is reduced indicating a greater probability of entrapment at local optima.
展开▼
