Study On Design And Application Of Artificial Immune Network Classifier

Artificial Immune System (AIS) is a new soft computing technology, and has been developed greatly in the last decade. Owing to the powerful information processing capability, AIS has been applied to spectrum areas, especially for the classification problems in the machine learning area. Among of the AIS models, the immune network can generalize the training antigen space quickly and effectively, and most of the developed immune based classifiers are designed and implemented with using immune network model. Although these classifiers have achieved great success for the application problems, they have some drawbacks, which limit their classification performance. The problems are listed as follows:1. There are no effective mechanisms used to guide the memory cells determination, so, the algorithms do not take the mutual relationship of the memory cells on the classification performance into account;2. There are no training space transformation methods used, the learning are usually conducted in the input spance directly, which limits the algorithm ability;3. Some AIS use linear mechanisms to control the antibody population evolution, which can not perturb the generation of antibody population, thus, the linear mechanism has negative effect on the sophisticated search capabilities of the algorithms;4. The memory cells are generated over randomly, and there are no effective methods presented to evaluate the memory cells and eliminate the bad cells;5. In the batch training process, there are no effective methods used to guide the antibody population evolution, the antibody population combination space is over large, which makes it very difficult to locate the optimal classifier.In order to solve the problems, this study proposed new methods to design artificial immune netwok classifiers. These new methods include pair wise antigens based training methodology, memory cell pruning method, kernel function and fuzzy logic. The details of algorithms are the followings.1. A new training method is proposed, this method uses pair wise antigens to guide the memory cells generation. For each training antigen, the algorithm determines the nearest antigen with different class with the training antigen as the pairing antigen of it. Then the candidate cell region is determined, which is a hyper sphere with the training antigen as the center and half of the distance between the2antigens as the radius. The training on the current antigen is finished when there are antibodies emerged in this region. Finally, the antibody in this region which is nearest to the pairing antigen is determined as the memory cell. This method not noly take the relationship of antigen-antibody into account, but also take the relationship between cells on the classification performance into account, which conducives to search optimal classifier. The algorithm has been used to6artificial datasets and5UCI datasets classification, and been applied to emotional speech recognition. The classification results are compared with the results obtained by other famous classifiers, such as SVM, C4.5, BayesNet, etc. The comparison shows that our algorithm has good classification performance.2. A kernel function based Artificial Immune Recognition System is proposed. In the conventional AIRS, the affinity of antigen-antibody is represented by Euclidean distance, which limits the algorithm nonliearity ability. Through using the kernel function, the input space is transformed into high dimensional feature space to improve the algorithm nonlinearity capacity. Moreover, the individual of the memory cell population is evaluated, and the weak cells which can not recognized any neighbor antigens correctly are removed. The algorithm has been5UCI datasets classification, and hepatitis and heart disease diagnosis, the diagnosis are evaluated by confusion matrix and AUC. From the classification performance comparison, it can been seen that the accuarcies reached by our algorithm are not only improved significantly when compared with the accuracies obtained by AIRS, but also higher than the accuracies reached by other classic classifiers.3. The classific AIRS uses linear resource allocation method, which can not perturb the antibody population evolution effectively and has negative effect on the sophisticated search capablities of the algorithm. In order to improve algorithm performance, there are2nonlinear resource allocation methods are presented. The2nonlinear methods are discrete resource allocation and fuzzy logic resource allocation. Discrete resource allocation method discretizes the whole stimulation interval into many sub intervals, and allocates equal resources to the antibodies in the same sub interval. Thus, the resource allocation can be changed only through optimizing the number of sub intervals without changing total resources, which perturbs the antibody population generation and improves the sophisticated search capablities of the algorithm. Fuzzy logic resource allocation uses a parameter to represent fuzzy logic, thus, the fuzzy logic for different problems need not to be pre-designed, the optimal fuzzy logic can be searched through only changing the parameter value, and thus the sophisticated search capabilities can be improved. Moreover, the memory cells are evaluated by fitness scores, and the cells of low fitness scores are pruned to improve the classifier. The algorithm has been used to6UCI classification, the results show the algorithm has good performance. Furthermore, the algorithm has been applied to heart disease, diabetes disease and breast cancer diagnosis, and the results are evaluated by AUC and confusion matrix. The results show that the algorithm is applicable to the3diseases diagnosis. Finally, the algorithm has been applied to bank user credit analysis, the comparison show that the algorithm has good performance for user credit analysis.4. Most of the developed immune network classifiers use incremental training method, although such method can optimize the memory cell for each antigen, it can not guarantee obtaining optimal memory cell population. In order to solve this problem, a tabu search strategy based immune network is proposed. The algorithm uses batch training methodology, such method presented the whole antigen population to the whole antibody population, and the classifier is evolved through evaluating the whohe antibody population. In order to reduce the search space effectively, the inner region of the class domains are forbiden to generated antibodies with the same classes as the inner antibodies, thus, the antibody generation with be mainly limited near the class domain boundary. Moreover, the antibodies are evaluated by the fitness scores and eliminate the antibodies of low fitness scores, which make the antibody population to generalize the training space effectively and represent the local characteristics of the training space. The algorithm has been used to4UCI datasets classification, the results show that the algorithm has excellent performance, especially for the Wine dataset, the accuracy reached by our algorithm is100%. Furthermore, the algorithm has been applied to emotional speech recognition and hepatitis and breast cancer diagnosis, the results show that the algorithm has good performance for these problems.