Improving Geometric Semantic Genetic Programming with Safe Tree Initialisation

Researchers in genetic programming (GP) are increasingly looking to semantic methods to increase the efficacy of search. Semantic methods aim to increase the likelihood that a structural change made in an individual will be correlated with a change in behaviour. Recent work has promoted the use of geometric semantic methods, where offspring are generated within a bounded interval of the parents' behavioural space. Extensions of this approach use random trees wrapped in logistic functions to parameterise the blending of parents. This paper identifies limitations in the logistic wrapper approach, and suggests an alternative approach based on safe initialisation using interval arithmetic to produce offspring. The proposed method demonstrates greater search performance than using a logistic wrapper approach, while maintaining an ability to produce offspring that exhibit good generalisation capabilities.