Improving the performance of evolutionary optimization by dynamically scaling the evaluation function

Traditional evolutionary optimization algorithms assume a static evaluation function, according to which solutions are evolved. Incremental evolution is an approach through which a dynamic evaluation function is scaled over time in order to improve the performance of evolutionary optimisation. We present empirical results that demonstrate the effectiveness of this approach for genetic programming. Using two domains, a two-agent pursuit-evasion game and the Tracker trail-following task (Jefferson et al., 1992), we demonstrate that incremental evolution is most successful when applied near the beginning of an evolutionary run. We also show that incremental evolution can be successful when the intermediate evaluation functions are more difficult than the target evaluation function, as well as when they are easier than the target function