Efficient Evolution of Neural Networks

Abstract

This thesis addresses the study of evolutionary methods for the synthesis of neural network controllers. Chapter 1 introduces the research area, reviews the state of the art, discusses promising research directions, and presents the two major scientific objectives of the thesis. The first objective, which is covered in Chapter 2, is to verify the efficacy of some of the most promising neuro-evolutionary methods proposed in the literature, including two new methods that I elaborated. This has been made by designing extended version of the double-pole balancing problem, which can be used to more properly benchmark alternative algorithms, by studying the effect of critical parameters, and by conducting several series of comparative experiments. The obtained results indicate that some methods perform better with respect to all the considered criteria, i.e. performance, robustness to environmental variations and capability to scale-up to more complex problems. The second objective, which is targeted in Chapter 3, consists in the design of a new hybrid algorithm that combines evolution and learning by demonstration. The combination of these two processes is appealing since it potentially allows the adaptive agent to exploit a richer training feedback constituted by both a scalar performance objective (reinforcement signal or fitness measure) and a detailed description of a suitable behaviour (demonstration). The proposed method has been successfully evaluated on two qualitatively different robotic problems. Chapter 4 summarizes the results obtained and describes the major contributions of the thesis.