Robustness of Artificially Evolved Robots: What's Beyond the Evolutionary Window? Abstract

Evolutionary methods are now commonly used to automatically generate autonomous controllers for physical robots. Although many evolutionary robotics studies have demonstrated the efficacy of this approach, few studies have actually focused on empirically testing the robustness of controllers synthesized using artificial evolutionary systems. In this paper, a self-adaptive Pareto evolutionary multi-objective optimization (EMO) algorithm called SPANN is used to evolve the locomotion controller of a virtual four-legged robot. The emphasis of this study is to observe the behavior of the artificial neural networks (ANNs) beyond the optimization window used to evolve the locomotion controllers. Specifically, the experiments will attempt to utilize the evolved controller beyond the time frame for which it is originally evolved to function as well a