Optimizing neural networks with genetic algorithms

We describe an approach to application-specific neural network design using genetic algorithms. A genetic algorithm is a robust optimization method particularly well suited for search spaces that are high-dimensional, discontinuous and noisy - features that typify the neural network design problem. Our approach is relevant to virtually all neural network applications: it is network-model independent and it permits optimization for arbitrary, user-defined criteria. We have developed an experimental system, NeuroGENESYS, and have conducted several experiments on small-scale problems. Performance improvements over manual designs have been observed, the interplay between performance criteria and network design aspects has been demonstrated, and general design principles have been uncovered.