Incremental Online Evolution and Adaptation of Neural Networks for Robot Control in Dynamic Environments

摘要

∗ These authors contributed equally. Abstract—Many approaches have been developed to tackle the design complexity of modern robotic systems by using evolutionary processes. Starting with an initial solution, the evolutionary process tries to adapt to a given scenario and in the end produces an improved solution. Previous work showed that incremental evolution, a stepwise increase in the scenario difficulty, can increase the success of evolutionary adaptation. In this work, we clearly confirm this effect in the context of online evolution of neural networks. The goal of our online evolutionary approach is to produce on average good, intermediate solutions while the system is adapting. We show that also the average performance of the continuous evaluations is increased by evolving first in a simple scenario and then transitioning to a more difficult scenario.