Backward-chaining genetic programming

Tournament selection is the most frequently used form of selection in Genetic Programming (GP). Tournament selection chooses individuals uniformly at random from the population. As noted in [6], even if this process is repeated many times in each generation, there is always a non-zero probability that some of the individuals in the population will not be involved in any tournament. In certain conditions, typical in GP, the number of individuals in this category can be large. Because these individuals have no influence on future generations, it is possible to avoid creating and evaluating them without altering in any significant way the course of a run. [6] proposed an algorithm, the backward chaining EA (BC-EA), to realised this, but provided limited empirical evidence as to the obtainable savings and experimentation was restricted to fixed-length genetic algorithms. In this paper we provide a genetic programming implementation of BC-EA and empirically investigate the efficiency in terms of fitness evaluations and memory use and effectiveness in terms of ability to solve problems of BC-GP. Our results indicate that the efficiency gains obtainable with this approach can be big