Online Evolution in Dynamic Environments using Neural Networks in Autonomous Robots

摘要

Abstract—Online evolution is adaptation of agents while they are deployed in their task. The agents adapt autonomously and continuously to changing environmental conditions and new challenges. Such changes are also a topic in incremental evolution, where the difficulty of a task is gradually increased in an attempt to increase adaptation success. Here we investigate an online evolutionary process in simulated swarm robots using recurrent neural networks as controllers. In order to cope with dynamic environments, we present a distributed online evolutionary algorithm that uses structural evolution and adaptive fitness. Using an experiment about incremental evolution as a test case, we show that our approach is capable of adapting to a change that requires new recurrent connections. Keywords-online evolution; incremental evolution; recurrent neural networks; swarm robotics; evolutionary robotics. I.