Evolutionary Robotics: Incremental Learning of Sequential Behavior

Abstract

Evolutionary robotics offers an efficient and easy-to-use framework for automatically building behaviors for an autonomous robot. However, a major drawback of this approach relies in the difficulty to define the fitness function (i.e. the learning setup) in order to get satisfying results. Recent works addressed this issue either by decomposing the learning task or by endowing the agent with such capabilities that should make the goal easier to achieve. Literature in evolutionary approach shows that modifying the very nature of genetic operators and/or fitness during the course of evolution may lead to better results for complex problems. In the scope of this short paper, we are interested in the reformulation of a straightforward complex fitness function into more subtle versions using different approaches