THE UNIVERSITY OF NEW SOUTH WALES SCHOOL OF COMPUTER SCIENCE AND ENGINEERING Evolving Neural Networks for Visual Processing

In this study a recently developed neuroevolution method, HyperNEAT, is applied to new tasks in order to learn more about its characteristics in general and for the tasks specifically. The tasks required directly processing raw image data, including some from a robot-mounted camera, requiring the neural networks to directly process hundreds or thousands of inputs. The tasks included object recognition and robot navigation. The complexity and properties of the tasks, as well as the neural network substrate topologies, were varied and the performance and characteristics of the evolutionary process and solutions generated were analysed. Positive results were achieved for non-trivial tasks, and some important characteristics of HyperNEAT not previously reported were discovered: a bias towards generating weight patterns aligned with the topographical arrangement of neurons within the layers of the substrate network; the impact on performance of the number of dimensions in the topographical arrangement of neurons in hidden layers; the ability to evolve solutions more quickly with larger substrate networks; and a possible difficulty co-evolving weight patterns for substrate networks containing more than one hidden layer. New directions for further research and possible enhancements to HyperNEAT are proposed based on these findings. Evolving Neural Networks for Visual Processing Oliver J. Coleman Supervisor: Alan Blair Pg. 2/69