Dynamical properties of the fitness landscape of a GP controlled random morphology robot

Abstract

The aim of this contribution is: (1) to present an easy to maintain robot hardware platform which allows on-line evolutionary experiments and demonstrations; (2) to introduce a simple method to measure dynamical characteristics of the time-dependent fitness landscape by using reference individuals; (3) to demonstrate dynamical properties of the fitness landscape based on fitness measurements of reference individuals. The implication of the observations for the design of on-line EAs in time-dependent fitness landscapes are discussed. Keywords: genetic programming, evolutionary robotics, on-line evolution, dynamical fitness landscape, reference fitness 1 INTRODUCTION In a conventional evolutionary algorithm (EA) [1] it is assumed that one fitness value is assigned to each individual 1 . In a typical implementation the fitness is evaluated, when the individual is created and stored together with the individual in the population data structure. In this case the fitness l...