The Immune Algorithm Based Classification Methods And Application

Artificial immune system (AIS) is a computing system for solving the complexproblems based on the function, principle, basic characteristics of biological immunesystem and relevant theories of immune theory. AIS combines some advantages ofclassifiers, neural network and machine reasoning system, could provide noiseendurance, self-organization, self-learning, memory and other evolutionary learningmechanism through natural defensive mechanism of learning techniques. AIS is a newintelligent computing research direction after artificial neural networks, evolutionarycomputation and the focus research of cross-disciplinary field with life sciences andcomputer science.In this thesis, the artificial immune algorithms are studied to solve classificationproblems of practical applications. Through the researches on immune mechanism ofnegative selection and clone selection, the multi-class classifier based on negativeselection is presented, and the improved clone selection algorithms are also introducedto solve the feature extraction problem and text classification problem. In addition, theintegrated theoretical framework between AIS and DNA computing is discussed, andDNA sequence encoding methods are studied. The main contributions and innovationsof this thesis are summarized as follows:①In order to solve the dimensionality reduction of high-dimension datasets in thefields of pattern recognition and information security, a feature selection approach basedon immune mechanism is designed to eliminate redundant attributes and reduce the sizeof the problem. This algorithm consists of three main stages: clonal selection, localsearch and classification. The antibodies with high affinity are obtained by clonalselection which has an inhibitory mechanism, then the local search operation is takenfor the optimal solution according to the weight of attribution for enhacing convergencerate, the selected subsets are tested and evaluated by K nearest neighbor classifier andclassification accuracies at last. The experimental results show that this algorithm couldreduce the feature dimension effectively and improve the classification accuracy.②The method for designing multi-class classifier which is based on negativeselection principle is proposed. The negative selection algorithm distinguishs theelements belonging to self from the elements belonging to nonself, in other word,classifies the problem with two categories. In this method, the self set and nonself set is redefined, sets of detectors are constructed for multi-class problem. The datasets withknown category are trained to generate detector set for each class to sort the datasetswith unknown category based on negative selection algorithm. The simulationexperiment on UCI standard data set proves the validity of the algorithm.③A clonal selection classification algorithm based on the antibody recognitionwhich considers antibody recognition ability to antigen as affinity function is introduced,that is, the affinity of antibody is evaluated by the number of correctly identifying theantigen with same class and the number of mistakenly identifying the antigen withdifferent class, then the optimal antibody is found by clone and mutation operation.During the searching period, the size of memory cells is controlled for reducing thenumber of memory cells while keeping the classification accuracy. Next, this algorithmis improved for applying to the web text classification. The most informative words areselected by information gain feature extraction method as feature vectors which arecalculated by TF-IDF formula. The datasets of feature vectors denoted for the web textare trained for memory cells of each class which will be used to classify the targetcollection for the category of target document.④Because the cloning and mutation operations occur randomly in many immunealgorithms, an immune algorithm based on particle swarm optimization is developed tosolve this randomness through swarm learning. The number of memory cells for eachcategory is predefined firstly to avoid the increase of antibodies in the training phase.The improved particle swarm optimization algorithm is used to evolve antibody. A newfitness function which considers the classification accuracy and distance betweenclasses is designed for evaluation of antibody affinity. In each iteration, each antibodyparticles moves to the optimal particle, in order to find the optimal solution quickly. Theexperiments on standard datasets confirm the validity of the algorithm.⑤The integrated theoretical framework of AIS and DNA computing is discussed.The difference between self and non-self can be done by string matching operation inartificial immune system, and the complementary base pairs can be done by sequenceoperation in DNA computing, so the feasibility of improving AIS algorithm by DNAcomputing is analyzed. After the research of molecular characteristics, a new DNAencoding method is proposed, which is based on the influence of molecular reactionprocess factors and can obtain reliable and valid coding sequence.The immune classification algorithms based on the negative selection and cloneselection are introduced in this thesis, which can be used to solve the classification problem in practical application.