Evolutionary Based Weight Decaying Method for Neural Network Training

摘要

A new weight decaying technique for neural network training is introduced. The proposed technique utilizes genetic algorithms in conjunction with a local optimization method to restrict the weights of the neural network in some range with desired generalization capabilities. This method is a global optimization one that overcomes most of the problems associated with local optimization procedures. In addition, this technique can be combined with any global optimization procedure from the relevant literature. The proposed technique has been evaluated on several well-known benchmark datasets, along with a series of classification and regression problems. The evaluation results are very promising indicating an improvement in classification error from 25 to 80% for the genetic algorithm.