Progress is not optimization : co-evolutionary robotics in simulation

Abstract

In this thesis, various aspects of the co-evolutionary projects of Nolfi and Floreano will be investigated. First of all, we will test the extent to which their "Evorobot" software allows one to replicate Nolfi and Floreano's experimental results. We will argue that numerous limitations and difficulties make it impossible for us to obtain complete resemblance. Sec- ondly, we will investigate whether progress throughout generations occurs. Moreover, we will test the progress for its monotonicity, using the isotonic regression analysis, and we will see that in all experiments the progress is significantly nondecreasing. Further, we will argue that evolving the individuals within more complex environments will lead to a larger amount of progress, while the use of environmental change or plasticity results in a smaller amount of progress. Finally, we will investigate the extent to which the co-evolutionary robotics of Nolfi and Floreano, as based on their software, is comparable to natural evolution. We will propose 3 criteria (i.e. the concepts, techniques used, and results), in order to examine the naturalness of their co-evolutionary robotics. We will claim that on the basis of these criteria there is no principled reason to deny the biological plausibility of Nolfi and Floreano's approach, although in practice, their co-evolutionary robotics is still far removed from it. Biological plausibility thus remains a fundamental important challenge for future research on co-evolutionary robotics.