The Optimization And Classification Algorithms Based On Artificial Immune System And Applications

Artificial Immune System (AIS) is a computing system that solves various complex problems based on the functions, disciplines, characteristics and other related immune theories of biological immune system. The system is a novel field of intelligent computing research after Artificial Neural Network (ANN) and Evolutionary Computation (EC). The objective of this study is to explore the evolutionary learning mechanisms contained in biological immune system and to apply them to the design of effective artificial immune models and algorithms for solutions of combination optimization problems in industries, and classification problems in Data Mining as well. The study focuses on the following aspects:1. The usual Clonal Selection Algorithm (CSA) is better than the Genetic Algorithm (GA) in global searching abilities, overcoming the problem of early-ripeness in function optimization. However, previous research demonstrates that when applying CSA to combination optimization problems such as 0-1 kidnap problems, it is incompetent of searching the best solution, slow of the processing speed of convergence and the solutions may fluctuate in a large scope. A Clonal Selection Algorithm with Receptor Editing (RECSA) is proposed, inspired by the biology immunity system mechanism. It makes use of not only cellular hypermutation but also receptor editing operation to realize affinity mature that will lead to the best match between the antibody and the antigen, Meanwhile a memory cell of the best individual descended is added to the mechanism to avoid population devolution. Aimed at the solution kidnap problems, the algorithm combines greedy strategy with an extended boundary to complete the receptor editing operation of each generation antibodies population. Applying RECSA to two 0-1 Kidnap Problems show that it can improve the population quality and search for the best solution more quickly than CSA. Its efficiency is higher and its stability and robustness is better than CSA and GA.2. Aimed at combinatorial optimization problems, a homogeneous finite Markov Chain Model using CSA and RECSA is constructed. The population state in Markov Chain is defined and a step transition probability matrix of system state is deduced. Then the convergence properties of the two algorithms are testified by the C-K equation properties in Markov chain theory. The testification process proved that the antibodies'population state can transit from arbitrarily beginning state to the best state when generation number is large enough, which means at least one best solution can be searched out in the final antibodies population. Lastly, testification based on the average absorbability time theorem in Markov Chain proves that the average generation number of convergence in RECSA is smaller than that in CSA, providing theoretical proof for the quicker convergence speed of RECSA in comparison with CSA.3. To show its universal validity of solving combined optimization problems, RECSA was further applied in the multicast routing field. In the case of multicast routing with delay constrain, the good gene segment in the immaturity subpopulation was adopted and receptor editing was conducted based on the principle of minimum cost and delay constrain separately. Results show that RECSA can search promptly for optimum solution without a prepared routing set. The process is also featured lower computational complexity and higher stability. RECSA has also been applied to the solution of the QoS problems in multicast routing. After fulfilling the condition of the time delay restraint, a parameter Q, which considers the balance the three performance including time delay, band width and the expense, is defined and introduced to measure the overall performance, maintaining the compromise and balance among the three Indies. This overcomes the problem in traditional solutions to multicast routing, namely the improvement of a performance parameter resulted in serious degeneration of other parameters. The results of simulation tests indicate its high searching efficiency, capable of adjusting the whole performance of the multicast routing and improved service quality of the QoS. At last, RECSA is applied to optimize LAC of Changsha mobile network so that number of times for LAC boundary locations update decrease while LAC sum is not changed.4. A classifier based on immune evolutionary network theories (IENC) has been proposed combining findings from studies on algorithms such as aiNET, AIRS and AINMC, and the immune evolutionary network theories. The classifier, based on the two network suppression operations within each memory cell pool and between memory cell pools, improves the network structure. Hence, the memory cellules can reach balance between peculiarity and universality, and eventually, improve the accuracy of classification. The results of tests on the four standard datasets of Iris, Ionosphere, Sonar and Pima in UCI indicate that IENC has better and higher accuracy than AIRS and AINMC .5. At last, satisfactory classification accuracy is achieved when applying IENC to DNA sequences and power quality disturbances detection. Euclid-distance was adopted for the affinity measurement of the classifier in all tests mentioned above. However, it was found that the eigenvalue extraction and the affinity measurement can negatively influence the classifier's characteristics in DNA sequence tests. IENC has therefore been improved as follows: the diversity increment was adopted to measure the affinity measurement, and a group of status parameters of the source of diversity are added to the respective frequency of the 64 codons of a sequence using sliding calculation, the improved classifier had resulted in better performance and higher accuracy. It is also more capable of judging similarities among sequences when applied to DNA sequences discrimination of three model species including C. elegans, S. cerevisiae and A. thaliana .