Evolving Artificial Neural Networks using Cartesian Genetic Programming

NeuroEvolution is the application of Evolutionary Algorithms to the training of Artificial Neural Networks. NeuroEvolution is thought to possess many benefits over traditional training methods including: the ability to train recurrent network structures, the capability to adapt network topology, being able to create heterogeneous networks of arbitrary transfer functions, and allowing application to reinforcement as well as supervised learning tasks. This thesis presents a series of rigorous empirical investigations into many of these perceived advantages of NeuroEvolution. In this work it is demonstrated that the ability to simultaneously adapt network topology along with connection weights represents a significant advantage of many NeuroEvolutionary methods. It is also demonstrated that the ability to create heterogeneous networks comprising a range of transfer functions represents a further significant advantage. This thesis also investigates many potential benefits and drawbacks of NeuroEvolution which have been largely overlooked in the literature. This includes the presence and role of genetic redundancy in NeuroEvolution's search and whether program bloat is a limitation. The investigations presented focus on the use of a recently developed NeuroEvolution method based on Cartesian Genetic Programming. This thesis extends Cartesian Genetic Programming such that it can represent recurrent program structures allowing for the creation of recurrent Artificial Neural Networks. Using this newly developed extension, Recurrent Cartesian Genetic Programming, and its application to Artificial Neural Networks, are demonstrated to be extremely competitive in the domain of series forecasting.