![]() | Only 14 pages are availabe for public view |
Abstract Research work in the area of extracting rules from tranined neural networks has observed much activity recently.Research work in the area of extracting rules from trained neural networks has observed much activity recently. However, the degree of complexity of ANN increases exponentially as a factor of the numbers of input and hidden nodes. The complexity problem can be improved by constructing the structure of the network based on constructive learning One of the most difficulties for the extraction of accurate knowledge is that the data being mined can be so noisy. In these cases neural networks are a feasible solution, due to their relatively good tolerance to noisy and generalization ability and the performance of a neural network is directly related to its parameters and architecture. But the difficulty is that the classification rules generated by neural networks are usually hardly comprehensible to the human users and this is a serious problem for the user. This is because neural networkcharacterized by real-valued parameters that are difficult to interpret. This problem can be solved by constructing hyper system from neural network and a new technique of Evolutionary Algorithms (EAs) called Gene Expression Programming (GEP).Gene expression programming (GEP) is the most recent technique of evolutionary algorithm, for data analysis. GEP uses fixed length, linear strings of chromosomes to represent computer programs in the form of expression trees of different shapes and sizes, and implements a GA to find the best program. Gene expression programming is one of several linear variations and liner chromosomes of GP and its aim is to improve the performance of GP and to offer many options of implementing GP in the 3rd and the 4th generations of computer programming languages.In this thesis, we presented destructive neural network learning technique and the analysis of the convergence rate of the error in a destructive neural network with and without threshold in the output layer at first.In addition, an algorithm has been proposed for rules extraction based on a trained neural network using Gene Expression Programming (GEP).We applied our method topublic-domain dataset in order to emphasize that the set of rules extraction from the proposed method is more accurate and brief compared with those obtained by the other models. |