A Research Based on a Modified Genetic Algorithm for the Overfitting of Resonance Maching Network

During the category learning process of SFAM ,due to the inputting category sample series containing noises or samples overlapping happened in any way easily, the phenomenon results to network structure redundancy and generalization ability problem.Aimed at it, this passage proposed a modified learning strategy based on an adaptive genetic algorithm. Firstly,to the variable structure genome sequence encoded composed of network conjunction weights and category nodes,we carry out a new operator—cutting operator,Instead of crossover operator in traditional GA, the operator depending on the rules of credibility judgement,cuts down the redundant category nodes which appeared in overfitting phenomenon. Secondly,to the mutation operator,we adjust the mutation rate by population fitness average. This passage takes the SFAM as a prototype, propose an modified learning strategy based on adaptive genetic al- gorithm which compared with SFAM and other improved ssFAM, ssGAM, ssEAM, Safe-μARTMAP network models,the experiment results show that ge- neralization ability has improved, and the modified network is obvious pro- moted in solving the overfitting phenomenon problem and reducing the network redundancy.