Genetic Programming with Historically Assessed Hardness

We present a variation of the genetic programming algorithm, called Historically Assessed Hardness (HAH), in which the fitness rewards for particular test cases are scaled in proportion to their relative difficulty as gauged by historical solution rates. The method is similar in many respects to some realizations of techniques such as implicit fitness sharing, stepwise adaptation of weights and fitness case selection, but the details differ and HAH is generally simpler and more efficient. It also leads to different generalizations. We present results from large-scale, systematic tests of HAH and we also discuss the technique in terms of the alternative explanations that it supports for the efficacy of implicit fitness sharing and related methods.