NeuroEvolution of Augmenting Topologies with Learning for Data Classification
Lin Chen, Damminda Alahakoon · 2006
Appropriate topology and connection weight are two very important properties a neural network must have in order to successfully perform data classification. In this paper, we propose a hybrid training scheme Learning-NEAT (L-NEAT) for data classification problem. L-NEAT simplifies evolution by dividing the complete problem domain into sub tasks and learn the sub tasks by incorporating back propagation rule into the NeuroEvolution of Augmenting Topologies (NEAT) algorithm. The new algorithm combines the strength of searching for topology and weights from NEAT and back propagation respectively while overcoming problems associated with direct use of NEAT. We claim that L-NEAT can produce neural network for classification problem effectively and efficiently. Empirical evaluation shows that L-NEAT evolves classifying neural network with good generalization ability. Its accuracy outperforms original NEAT.