A Relaxed Approach to Simplification in Genetic Programming

We propose a novel approach to program simplification in tree-based Genetic Programming, based upon numerical relaxations of algebraic rules. We also separate proposal of simplifications from an acceptance criterion that checks the effect of proposed simplifications on the evaluation of training examples, looking several levels up the tree. We test our simplification method on three classification datasets and conclude that the success of linear regression is dataset dependent, that looking further up the tree can catch ineffective simplifications, and that CPU time can be significantly reduced while maintaining classification accuracy on unseen examples.