Evolutionary Design of Neural Networks

This thesis deals with methods for nding neural network architectures suitable for learning particular problems. We use an evolutionary algorithm with four di erent genetic encoding methods to search for the suitable architectures. We train the neural network weights with a separate neural learning algorithm. We use eight di erent learning problems for benchmarking the encoding methods. Four of the problems are arti cial (XOR, Encoder and two function approximation problems), three are real-world classi cation problems from the Proben1 benchmarking problem set, and one is a bankruptcy classi cation problem studied earlier in one of our projects. Our evaluation criteria are classi cation accuracy and e ciency for using only the relevant variables. The classi cation results are compared also to those for network architectures found by a systematic search.