Genetic programming of logic-based neural networks

Genetic algorithms and genetic programming are optimization methods in which potential solutions evolve via operators such as selection, crossover and mutation. Logic-Based Neural Networks are a variation of artificial neural networks which fill the gap between distributed, unstructured neural networks and symbolic programming. In this thesis, the Genetic Programming Paradigm is modified in order to obtain Logic-Based Neural Networks. Modifications include connection weights on the parse trees, a new mutation operator, a new crossover operator, and a new method for randomly generating individuals. The algorithm is part of a two-level development process where, at first, satisfactory logic-based neural networks are obtained using our algorithm; then, gradient-based learning methods are used to refine the networks. Results are obtained for a 6-input Logic-Based Neural Network problem.