Age based controller stabilization in Evolutionary Robotics

Abstract

Evolutionary Robotics is a collection of heuristics where robotic control systems are developed by following the example of natural evolution. An evolutionary run is performed by mutating the robots' controllers randomly and selecting for some desired behavioral properties. Overall, these properties should be improved over time leading to a stable increase of fitness. However, random mutations on critical controller parts can lead to a rapid degradation lowering the performance of evolution. This paper presents an approach to reduce the loss of desirable behavior during an evolution process. A notion of age is introduced as a quality criterion to indicate the contribution of parts of a controller to the robot's overall behavior. To preserve the behavior evolved so far, mutations are channeled to affect controller parts with a lower age more than those with a higher age. As a result, controller parts that contribute to a good behavior are stabilized and the evolved desirable behavior is maintained. Experiments have been performed in a decentralized online evolutionary scenario with controllers based on finite state machines (FSMs). The results show an improvement in the number of successful evolutions and the number of successfully evolved robots compared to previous studies.