Font Size: a A A

Research On Key Issues For Attribute Co-evolutionary Reduction In Rough Set Theory

Posted on:2014-10-20Degree:DoctorType:Dissertation
Country:ChinaCandidate:W P DingFull Text:PDF
GTID:1268330422980008Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Rough set theory (RST) was introduced by Professor Pawlak in1982to analyze and deal withsome information and knowledge with imprecision, inconsistency and incompletion. In recent years,it has been widely applied in diversified research areas, such as data mining, pattern recognition,machine learning, and so on. Attribute reduction is one of the most important topics of RST, and it hasbeen recognized as an important feature selection method. Attribute reduction studies how to removeirrelevant and redundant features with minimal information loss, and to select the attribute subsetfrom original attributes in the decision table while retaining the suitably high classificationperformance in representing the original attributes. Finding the minimum reduction set is moredifficult and it has been proven to be a representative NP-hard problem by Wong et al, and manyresearchers have made efforts on the algorithms of attribute reduction and achieved some betterresults. However, there has been no universal and efficient solution for this purpose.In recent years, the co-evolutionary algorithm has been a growing interest in the field ofcomputational intelligence by revealing and simulating the co-evolutionary phenomenon and process ofmulti-populations in the natural ecosystem. So far it has been proven to be effective in solving manycomplex problems of which are difficult by using the traditional evolutionary algorithms, especially forsome NP-Hard problems. In this thesis, the co-evolutionary algorithm is introduced into theoptimization problem of minimum attribute reduction in RST. Some key problems for attributeco-evolutionary reduction such as convergence, cooperation mechanism, model optimization,evolutionary adaptability, large-scale attribute reduction and its representative individual selection arethoroughly investigated, so that theoretical work of attribute reduction in RST is perfected. A series ofmodels and algorithms about attribute reduction are developed under the co-evolutionary frameworkand applied into some practical applications, such as attribute reduction and feature classification,magnetic resonance images (MRI) reduction and segmentation, and so on. Some experimental resultsdemonstrate the effectiveness of proposed algorithms. In this thesis, some key problems on attributeco-evolutionary reduction in RST are systematically and deeply researched, in which the traditionalalgorithms are improved, and some new models and algorithms are put forward. The related researchresults significantly enhanced the performance of attribute co-evolutionary reduction.More concretely, the main contributions of this thesis include:1. For the problems on traditional attribute evolutionary reduction in the premature convergenceand population neighborhood selection, the niche technology is introduced into the attributeevolutionary reduction, and a niche conic neighborhood particle swarm optimization based attribute co-evolutionary reduction algorithm is proposed. Particle swarm with social cognition behavior ismapped into the attribute approximate space for iterative optimization. It can self-adaptively constructthe niche vector of neighborhood by the layered conic space. And the adaptive penalty function isused to strengthen the convergence ability of population fitness, so that the premature convergencecan be avoided well and the minimum attribute reduction set is obtained quickly. Experimental resultsdemonstrated that performance of proposed algorithm is improved much better, by playing the effectof populations’ co-evolutionary reduction within their respective niche neighborhoods as to avoid thepremature convergence.2. In order to improve the co-evolutionary performance for attribute reduction in RST, a noveland efficient self-adaptive evolutionary tree based attribute mixed co-evolutionary reductionalgorithm is put forward by expanding the essence of co-evolutionary mechanism. It constructs a kindof the dynamic population model with the self-adaptive multilevel evolutionary tree. The mixedco-evolutionary mechanism of competitive and cooperative co-evolution is adopted to select therespective excellent individual in each evolutionary sub-tree, and to carry on their attribute subsetvector to share the co-evolutionary searching experience. The proposed algorithm can better balancebetween exploitation in breadth and exploration in depth. Some experiments are presented to show theeffectiveness of the mixed co-evolutionary mechanism with competitive and cooperativeco-evolution.3. The quantum-inspired co-evolutionary algorithm is firstly introduced into the research for themodel optimization of attribute reduction in RST, therefore an efficient quantum frog based attributeco-evolutionary reduction algorithm is proposed. In this algorithm, individuals with dynamicmulti-cluster frog structure are represented by multi-state quantum bits, in order to increase thediversity of evolutionary individuals. The self-adaptive adjustment of quantum rotation gate speedsthe search process so that it can keep balance between global search and local refinement during theattribute co-evolutionary reduction. The strategies of quantum mutation and quantum entanglementare applied to accelerate the evolution convergence. The algorithm reconstructs the fitness function ofattribute reduction optimization based on quantum co-evolution, in order to obtain more satisfyingminimum attribute reduction set. And on such basis, the attribute co-evolutionary reduction andclassification learning cascade algorithm based on the quantum frog with adaptive crossovercooperation is proposed so as to further improve the performance of attribute reduction andclassification learning for decision rules. Experimental results indicate it has extremely strong globaloptimization of attribute co-evolutionary reduction and has achieved better high-performance onefficiency and accuracy than traditional algorithms. From the perspective of quantum co-evolution,attribute quantum co-evolutionary reduction algorithm is designed, which will provide a better idea for new framework of attribute evolutionary reduction.4. In order to further improve the adaptability of attribute quantum co-evolutionary reduction, aquantum cloud model based attribute self-adaptive co-evolution reduction algorithm is presented,according to outstanding characteristics of the cloud model on the process of transforming aqualitative concept to a set of quantitative numerical values. First, quantum population gene cloud isused to encode the evolutionary frog population, and reversible cloud mode based on attribute entropyweight is designed to adjust the quantum revolving gate adaptively, so the scope of search space canbe adaptively controlled under the guidance of qualitative knowledge. Second, both the quantumcloud mutation and quantum cloud entanglement operators are used to make quantum frog populationbe adaptive to get the optimization attribute reduction sets fast. By using of cloud model, the proposedalgorithm can make attribute quantum co-evolutionary reduction with stronger adaptability and canbetter deal with the vague and incomplete attribute reduction.5. Aiming at the optimization problem of application in large-scale attribute reduction and itsrepresentative elitist individual selection, a novel quantum elitist frog based large-scale attributeco-evolutionary ensemble reduction algorithm is proposed by enlightening the elite role attitude underthe collective collaboration model. First, the algorithm designs the multilevel elitist pool of quantumfrogs, in which quantum elitist frogs can fast guide the evolutionary population into the optimal area.Second, a self-adaptive cooperative co-evolutionary model is constructed to decompose thelarge-scale attribute set into reasonable-scale attribute subsets according to the best historicalperformance experience records and assignment credits. Third, some optimal elitists in differentmemeplexs are selected out to evolve their representing attribute subsets, which can increase thecooperation and efficiency of large-scale attribute reduction. Therefore the global minimum attributereduction set can be gained steadily and efficiently. Theoretical analysis and experimental results arepresented to show the feasibility and effectiveness of the proposed algorithm. Finally this proposedalgorithm is applied into magnetic resonance images (MRI) reduction and segmentation, and theeffective and robust segmentation results further demonstrate it has stronger applicability.
Keywords/Search Tags:Attributes reduction, co-evolutionary algorithm, niche conic neighborhood, self-adaptive evolutionary tree, shuffled frog-leaping algorithm, quantumevolutionary operator, quantum cloud model, quantum population elite
PDF Full Text Request
Related items