Experiments on Neuroevolution and Online Weight Adaptation in Complex Environments

Neuroevolution has come a long way over the last decade. Lots of interesting and successful new methods and algorithms have been presented, with great improvements that make the field become very promising. Concretely, HyperNEAT has shown a great potential for evolving large scale neural networks, by discovering geometric regularities, thus being suitable for evolving complex controllers. However, once training phase has finished, evolved neural networks stay fixed and learning/adaptation does not happen anymore. A few methods have been proposed to address this concern, mainly using Hebbian plasticity and/or Compositional Pattern Producing Networks (CPPNs) like in Adaptive HyperNEAT. This methods have been tested in simple environments to isolate the effectiveness of adaptation from the Neuroevolution. In spite of this being quite convenient, more research is needed to better understand online adaptation in more complex environments. This paper shows a new proposal for online weight adaptation in neuroevolved artificial neural networks, and presents the results of several experiments carried out in a race simulation environment.