Abstarct: In the last decade the use of artificial intelligence techniques to solve scientific problems has increased dramatically. Artificial neural networks in classification and pattern recognition issues are widely used. The lack of capability in interpreting the results is one of the biggest problems of artificial neural networks. Although, artificial neural networks are able to carry out the classification operations accurately the lack of transparency of the knowledge provided by the neural network makes the results can’t be easily used in other systems. Therefore, researchers are trying to fix this problem by extracting the rules of neural networks. Today, there are several methods of rule extraction which are often able to present the operations performed by an artificial neural network on a simple data collection in the form of a set of simple rules. However, these methods are usually less accurate in the case of relatively complex issues and the precision resulted from rules is not comparable to the results of the neural network. In this project, a new method for extracting the rule from artificial neural networks tixtled as Four Step Rule Extraction (FSRE) has been introduced. This method is able to convert the classification carried out on a data set into a set of simple, useful and efficient propositional rules. In using this method, there is no limitation on the complexity of the data. To evaluate the performance and results of this method, the common classification data sets (from the databaxse UCI) is used in this thesis. The results showed that the proposed method is more efficient and accurate compared to the available methods of the rule extraction.