An investigation of dynamic fitness measures for genetic programming

This research investigates the hypothesis that the use of different fitness measures at the different generations of genetic programming (GP) is more effective than the convention of applying the same fitness measure individually throughout GP. A genetic algorithm (GA) is used to induce the sequence in which fitness measures should be applied over the GP generations. Subsequently, the performance of a GP system applying the evolved fitness measure sequence is compared with the conventional GP approach. The former approach is shown to significantly outperform standard GP on varied benchmark problems. Furthermore, the evolved fitness measure sequences are shown to generalize within a problem class: therefore, the sequences can be evolved off-line for different problem classes. Critically, sequences trained on the problem classes are also shown to generalize to complex, real-world problems. Overall, the findings of the study are in favor of the hypothesis. This study has revealed the effectiveness of dynamic fitness measures when applied to benchmark and real-world problems.