Evolving developmental, recurrent and convolutional neural networks for deliberate motion planning in sparse reward tasks

Abstract

Motion planning algorithms have seen a diverse set of approaches in a variety of disciplines. In the domain of artificial evolutionary systems, motion planning has been included in models to achieve sophisticated deliberate behaviours. These algorithms rely on fixed rules or little evolutionary influence which compels behaviours to conform within those specific policies, rather than allowing the model to establish its own specialised behaviour. In order to further these models, the constraints imposed by planning algorithms must be removed to grant greater evolutionary control over behaviours. That is the focus of this thesis.
An examination of prevailing neuroevolution methods led to the use of two distinct approaches, NEAT and HyperNEAT. Both were used to gain an understanding of the components necessary to create neuroevolution planning. The findings accumulated in the formation of a novel convolutional neural network architecture with a recurrent convolution process. The architecture’s goal was to iteratively disperse local activations to greater regions of the feature space. Experimentation showed significantly improved robustness over contemporary neuroevolution techniques as well as an efficiency increase over a static rule set. Greater evolutionary responsibility is given to the model with multiple network combinations; all of which continually demonstrated the necessary behaviours. In comparison, these behaviours were shown to be difficult to achieve in a state-of-the-art deep convolutional network.
Finally, the unique use of recurrent convolution is relocated to a larger convolutional architecture on an established benchmarking platform. Performance improvements are seen on a number of domains which illustrates that this recurrent mechanism can be exploited in alternative areas outside of planning. By presenting a viable neuroevolution method for motion planning a potential emerges for further systems to adopt and examine the capability of this work in prospective domains, as well as further avenues of experimentation in convolutional architectures.