Research On Some Key Issues In Artificial Immune Systems

Artificial immune system, inspired by natural immune principle, is a new computational model, and it contains three computational models:negative selection algorithm, clonal selection algorithm, and immune network. Negative selection algorithm simulates T-cell's abilities of identifying self and non-self accurately, and is applied to generate abnormal detection systems. But systems exists holes so that some non-self individuals are not detected. On the other hand, the computational complexity of generating abnormal detection systems is high. The above aspects of negative selection algorithm affect its applications in real world. The second computational model is clonal selection algorithm. The algorithm simulates the process of immune response triggered when natural immune system is attacked by external entities. Clonal selection algorithm is employed to global high-dimensional optimization problems. However, the algorithm demonstrates premature convergence and insufficient of diversity information. The next computational model is immune network, which tries to simulate the whole natural immune network. The model which contains many control parameters and few attention is focused on it.In the paper, some problems coming from negative selection algorithm and clonal selection algorithm are studied. Main contributions and innovations of the dissertation are shown as follows.(1) A new detector generation strategy, based on seed individuals and contiguous somatic simulating mutation, was proposed to reduce the time complexity of negative selection algorithm (NSA). Firstly, the strategy produced seed detectors, and determined the special detectors and gene segment by measuring the affinity between seed set and self set; Then a stimulated-response mutation (SRM) occurred in a special gene fragment to obtain the newly candidate individuals; Finally the new competent detectors were selected according to r-contiguous bits matching rule. The characteristic of the algorithm is the pattern information used to guide the mutation process for reducing the matching rate of candidate individuals. The experimental results show that the algorithm outperforms several similar algorithms based on mutation operator in term of time complexity and coverage. (2)Hierarchy Match Strategy (HMS) is proposed to decrease the computational cost of negative selection algorithm. The foundation of HMS is proved in theoretical firstly. Then HMS constructs the components of detector set through dividing the self set into several pattern subsets according to r-contiguous bit match rule. Lastly Detector set is obtained by using binary tree combing components. Using the self pattern information in generating process is the novelty of the algorithm, and which is the difference between the conventional generation strategies. The experimental result shows that HMS improves the performance in term of both detective rate and time complexity under the same experimental environment.(3) A novel exploring holes algorithm based on non-detector pattern (short for EHANDP) was proposed for holes existing in anomaly detection system generated by negative selection algorithm. Incompleteness of current exploring holes algorithm grounded on self pattern (short for EHASP) was point out. And then the sufficient and necessary condition for individuals to be holes was proven using the string patterns in problem space, what is more, an exploring holes algorithm named EHANDP was proposed. The capability of finding all holes of a given detection system is EHANDP's main feature. The above two algorithms are compared using random dataset and artificial dataset, and the results shows that EHANDP algorithm outperforms than EHASP in the term of exploring capability although they have the same computational complexity.(4) In order to overcome the premature of immune algorithm when solving high dimensional multimodal functions, an efficient hybrid immune evolutionary algorithm is proposed. The main characteristics of the novel hybrid algorithm are dynamic clonal selection, archive-based hypermutation and multi-parentic crossing operators. In addition, a novel performance evaluation criterion for comparing different algorithms is constructed in the paper. In experimental study, firstly the performance of proposed HIEA is validated using several classical test functions; next HIEA is compared with self-adaptive differential evolution (SaDE) and simple immune algorithm (SIA) under certain amount of function evaluations, the experimental results show that the performance of proposed HIEA is significant better than that of SaDE and SIA in term of the accuracy and stability. (5) In order to increase the diversity of immune algorithm when solving high dimensional global optimization problems, a novel immune evolutionary algorithm (IEA) is proposed. The main characteristics of IEA are clonal expansion and multiple-parent random receptor editor operators. In addition, a modified hypermutation operator is introduced to improve the learning ability of individuals. In the experimental study, firstly several typical test functions are used to determine the population size and the ratio of clonal expansion. Next, the IEA is compared with fast clonal algorithm (FCA) and Opt-IMMALG, and the experimental results of the IEA are significantly better than that of FCA and Opt-IMMALG in terms of the performance evaluation criterion proposed.