Research On Some Issues Of Quantum-behaved Particle Swarm Optimization Algorithm And Its Applications

Evolutionary computation is a kind of artificial intelligence technology, which solves theoptimization problem through simulating the biological evolutionary process and mechanismsin nature. It starts from random solutions, obtaining the optimal solution by the iterativeprocess. In the iterative process, the fitness function is adopted to evaluate the solution. Somerepresenting evolutionary algorithms are listed as follows: Genetic Algorithm, EvolutionaryStrategies, Evolution Programming, Genetic Programming, Particle Swarm Optimization(PSO), which is inspired by bird flock looking for food, and Quantum-behaved ParticleSwarm Optimization (QPSO). PSO algorithm and QPSO algorithm are new evolutionarycomputation technique developed in recent years. In PSO algorithm, each potential solution ofoptimization algorithm, call a bird, flies in the search space. PSO algorithm is characterizedby simplicity in parameters setting, quickness in computing speed, and excellence in localsearch performance. However, the algorithm can not converge to the global minimum pointwith probability one under suitable condition. Base on the deep study of PSO algorithm andinspired by quantum physics, Quantum-behaved Particle Swarm Optimization (QPSO)algorithm is proposed. QPSO algorithm has much fewer parameters without the velocity ofparticles and much stronger global search ability than the PSO algorithm.Theoretical analyses and algorithm improving on PSO algorithm and QPSO algorithmare mainly discussed in the dissertation and the application of QPSO algorithm on fuzzyneural networks (FNN), Bayesian network and protein folding are also studied. The maincontents are as follows:(1) Firstly, the background of the research work, including the basic concept of Bayesiannetwork learning, fuzzy neural networks, protein folding and the development of someevolutionary algorithms are introduced. Then the achieved results and future research isreviewed. On this basis, the research topic and significance of this subject are presented. Thebasic principles and models, the behavior analysis and convergence analysis of PSO algorithmand QPSO algorithm are described in detail. Finally some improved algorithms are proposed.(2) The major drawback of QPSO algorithm and some improved QPSO algorithms isthat the dimension of search space is fixed. In many optimization problems, the optimumdimension is dynamic. In order to address this problem, the variable-dimensional QPSO(VDQPSO) algorithm, which negates the need of fixing the dimension of the solution space inadvance, is presented. In the proposed algorithm, the optimal dimension, which is regarded asan optimization objective, is updated together with the position of particles. At last, all theparticles can converge to the global solution on the optimum dimension in a simultaneous way.The structure learning of FNN, which is a NP-hard problem, is adopted to test theperformance of VDQPSO. It is difficult to determine the proper number of fuzzy rules inadvance. Most algorithms obtain the number of fuzzy rules by expert experiences. The noveltechnique can address the drawback. And the optimum dimension converged at the end of theoptimization process corresponds to a unique FNN structure where the optimum parameterscan be achieved. (3) A novel QPSO algorithm with comprehensive learning and cooperative learningapproach (CCQPSO) is introduced to improve the premature convergence of QPSO forsolving multimodal problems. The comprehensive learning strategy of algorithm can makethe swarm diversity increase and improve the global convergence property of QPSO. Thecooperative learning approach can avoid loss some components that had moved closer to theglobal optimal solution in the vector by feeding back each dimension update of particle, andlead the algorithm to local search quickly. In the proposed algorithm, the position of localattractors is updated by comprehensive learning strategy firstly. Then splits the particlesolution with composite high-dimensional into several one-dimensional sub-parts andevaluates each updated dimension of particle solution through adopting a crossover operationthat similar to GA. The results of standard test functions experiment and learning of FNNhave showed that in contrast to other algorithm, the CCQPSO algorithm performs better onglobal convergence and has stronger ability to escape from the local optimal solution duringthe search process especially with high dimension multimodal functions.(4) Bayesian network is a language to describe the relationship systematically betweenrandom variables. The conformation of Bayesian network includes two way: one is constructmanually by expert experiences, the other is obtain the network by analyzing data set withmachine learning. The former method is subjective. Hence the latter method is the ourresearch topic. The conformation process of Bayesian network is an optimization problem sothat it can be optimize by discrete QPSO algorithm. Comprehensive learning and cooperativelearning approach are introduced too to address the premature convergence appeared indiscrete QPSO algorithm. The proposed approach is used in structure learning of Bayesiannetwork, which is represented by adjacent matrix or vectors. Unreasonable structure iseliminated by mutual information between the variables in the process. After that the optimalnetwork can be obtained. For evaluating the best matching degree between Bayesian networkand sample data sets, Bayesian information criterion(BIC) score is proposed. Finally, twobenchmarks of Bayesian networks are used to test the new approach. The results ofexperiments show that the proposed technique converges more rapidly than other evolutionarycomputation methods.(5) Protein folding problem is to predict the dimensional folding configurations from itsamino acid sequence. The HP lattice model, which depicts the protein, and the globalminimum energy conformation, which is fitness function, are introduced in this paper. Theestimation of distribution algorithms based on statistical learning and probability distributionmodel are introduced to discrete QPSO algorithm for protein folding. The improved algorithmhas the advantage of estimation of distribution algorithms and discrete QPSO algorithm,which includes the global convergence property of the estimation of distribution algorithmsand the local convergence property in the latter iterative of the discrete QPSO algorithm. Thenovel algorithm can not only locate quickly the approximate range of the optimal solution, but also converge rapidly to the optimal solution. Nine amino acid sequences are suggested to testthe performance of improved algorithm.