Lifetime Adaptation in Genetic Programming for the Symbolic Regression

Abstract

This paper focuses on the use of hybrid genetic programming for the supervised machine learning method called symbolic regression. While the basic version of GP symbolic regression optimizes both the model structure and its parameters, the hybrid version can use genetic programming to find the model structure. Consequently, local learning is used to tune model parameters. Such tuning of parameters represents the lifetime adaptation of individuals. Choice of local learning method can accelerate the evolution, but it also has its disadvantages in the form of additional costs. Strong local learning can inhibit the evolutionary search for the optimal genotype due to the hiding effect, in which the fitness of the individual only slightly depends on his inherited genes. This paper aims to compare the Lamarckian and Baldwinian approaches to the lifetime adaptation of individuals and their influence on the rate of evolution in the search for function, which fits the given input-output data.