Abstarct: Nowadays with the growing volume of data, analyzing and discovering the relationships between them, has become very important in data mining science. Rule Extraction is one of the important applications for discovering knowledge and relationships between data. It is done by investigating the internal structure of the network, connections and the output of the hidden laxyer neurons. Neural networks, especially, have received a lot of attention because of their high accuracy and not requiring prior knowledge about data. Most algorithms in this field of study use decompositional approach and are dependent on pruning phase. However, pruning reduces the efficiency and accuracy of a network which lead to weakening the extracted rules. We proposed a new method to reduce pruning damage using genetic algorithm at two general and local levels. In this method, neural network weights are adjusted in a way, so that the damage caused by the pruning phase will be reduced significantly. The effectiveness of the proposed approach is clearly demonstrated by the experimental results on a set of standard data mining test problems. The accuracy of the rules extracted by our proposed method is improved about 2 percent on average.