[PDF] [BIB] [PUB]
The optimization of the control parameters of genetic algorithms is often a time consuming and tedious task. In this work we take the meta-level genetic algorithm approach to control parameter optimization. We enhance this process by incorporating a neural network for fitness evaluation. This neural network is trained to learn the complex interactions of the genetic algorithm control parameters and is used to predict the performance of the genetic algorithm relative to values of these control parameters. To validate our approach we describe a genetic algorithm for the largest common subgraph problem that we develop using this neural network enhanced meta-level genetic algorithm. The resulting genetic algorithm significantly outperforms a hand-tuned variant and is shown to be competitive with a hill-climbing algorithm used in practical applications.