Fitness Sharing in Genetic Programming

摘要

This paper investigates fitness sharing in genetic programming. Implicit fitness sharing is applied to populations of programs. Three treatments are compared: raw fitness, pure fitness sharing, and a gradual change from fitness sharing to raw fitness. The 6- and 11-multiplexer problems are compared. Using the same population sizes, fitness sharing shows a large improvement in the error rate for both problems. Further experiments compare the treatments on learning recursive list membership functions; again, there are dramatic improvements in error rate. Conversely, fitness sharing runs achieve comparable results to raw fitness using populations two to three times smaller. Measures of population diversity suggest that the results are due to preservation of diversity and avoidance of premature convergence by the fitness sharing runs.