Modelling evolvability in genetic programming

Abstract

We develop a tree-based genetic programming system, capable of modelling evolvability during evolution through artificial neural networks (ANN) and exploiting those networks to increase the generational fitness of the system. This thesis is empirically focused; we study the effects of evolvability selection under varying conditions to demonstrate the effectiveness of evolvability selection. Evolvability is the capacity of an individual to improve its future fitness. In genetic programming (GP), we typically measure how well a program performs a given task at its current capacity only. We improve upon GP by directly selecting for evolvability. We construct a system, Sample-Evolvability Genetic Programming (SEGP), that estimates the true evolvability of a program by conducting a limited number of evolvability samples. Evolvability is sampled by conducting a number of genetic operations upon a program and comparing the fitnesses of resulting programs with the original. SEGP is able to achieve an increase in fitness at a cost of increased computational complexity. We then construct a system which improves upon SEGP, Model-Evolvability Genetic Programming (MEGP), that models the true evolvability of a program by training an ANN to predict its evolvability. MEGP reduces the computational cost of sampling evolvability while maintaining the fitness gains. MEGP is empirically shown to improve generational fitness for a streaming domain, in exchange for an upfront increase in computational time.