Investigating the effect of pruning on the diversity and fitness of robot controllers based on MDL2∈ during Genetic Programming

In this paper we propose a new diversity measure based on the correlation of bit strings for the analysis of genetic programming (GP) experiments. The diversity measure has been applied to analyse the impact of pruning on the diversity of a population during genetic programming and its relation to the convergence time of the fitness function. To show the usability of the proposed diversity measure a GP experiment is introduced where simulated Jasmine robots have to learn a collision avoidance behaviour to find their way through a maze. A full analysis of this experiment is given with different fixed pruning strategies in respect to the population diversity and fitness. The GP has been done on behaviour-based robot controllers implemented in MDL2isin. MDL2isin has the advantage that it provides a very compact bit string representation of the control programme, which can be used for diversity analysis.