Evolution of neural networks for physically simulated evolved virtual quadruped creatures

This work develops evolved virtual creatures (EVCs) using neuroevolution as the controller for movement and decisions within a 3D physics simulated environment. Previous work on EVCs has displayed various behaviour such as following a light source. This work is focused on complexifying the range of behaviours available to EVCs. This work uses neuroevolution for learning specific actions combined with other controllers for making higher level decisions about which action to take in a given scenario. Results include analysis of performance of the EVCs in simulated physics environment. Various controllers are compared including a hard coded benchmark, a fixed topology feed forward artificial neural network and an evolving ANN subjected to neuroevolution by applying mutations in both topology and weights. The findings showed that both fixed topology ANNs and neuroevolution did successfully control the evolved virtual creatures in the distance travelling task.