ABSTRACT
When the ordinary backpropagation neural network is trained, the problem of seriously local minimum often occurs. This makes net-training unfinished and weights distributed in the network immature. The current paper proposes a strategy of optimizing neural networks by genetic algorithm to deal with this problem. If the initially decided conditions (including training time and error precision) cannot be satisfied, i.e., the neural networks settle in local minimum, the unaccomplished network turns to be optimized by the genetic algorithm through adjusting the immature weights. Then the evolutionary weights which are believed to grasp the property of the sample construct a new network model which shows strong discriminant power. By use of the proposed evolutionary strategy, a credit scoring model applied to Chinese business corporations is developed and the scoring power of the model is tested in this study. Finally conclusions are presented.
