Committee learning of partial functions in fitness-shared genetic programming

This paper investigates the application of committee learning to fitness-shared genetic programming. Committee learning is applied to populations of either partial and total functions, and using either fitness sharing or raw fitness, giving four treatments in all. The approaches are compared on three problems, the 6- and 11-multiplexer problems, and learning recursive list membership functions. As expected, fitness sharing gave better performance on all problems than raw fitness. The comparison between populations of partial and total functions with fitness sharing is more equivocal. The results are very similar, though slightly in favour of total functions. However there are strong indications that the average size of individuals in the partial function populations are smaller, and hence might be expected to generalise better, though this was not investigated in this paper