Central Point Crossover for Neuro-genetic Hybrids

In this paper, we consider each neural network as a point in a multi-dimensional problem space and suggest a crossover that locates the central point of a number of neural networks. By this, genetic algorithms can spend more time around attractive areas. We also apply representational normalization to neural networks to maintain genotype consistency in crossover. For the normalization, we utilize the Hungarian method of matching problems. The experimental results of our neurogenetic algorithm overall showed better performance over the traditional multi-start heuristic and the genetic algorithm with a traditional crossover. These results are evidence that it is attractive to exploit central areas of local optima.

• Type: Conference
• Year: 2004
• Where: GECCO
• Authors: Soonchul Jung, Byung Ro Moon
• Keywords: GECCO 2004, Genetic Algorithm, Multi-dimensional Problem Space, Neural Networks