Research On Multi_Objective Optimization Algorithm For Biclustering In Microarry Gene Expression Data

Modern microarray technologies can generate massive amounts of geneexpression datasets that contain expression levels of thousands of genes underhundreds of different experimental conditions. In order to discover novel and usefulknowledge from those datasets more effectively and efficiently, there is an increasingneed to develop more powerful data mining algorithms. Biclustering, which cancluster both microarray gene and conditions simultaneously, is a very practical datamining technique in microarray gene expression data analysis. The objective ofbiclustering is to find maximal sub-matrices where the genes exhibit highly correlatedactivities over a subset of conditions. Because there are several mutually conflictinggoals, multi-objective optimization algorithms are good candidate solutions.Nature-inspired algorithms, such as particle swarm optimization (PSO) andsimulated annealing (SA), are metaheuristics that mimics the nature for solvingoptimization problems. They have been successfully used in many multi-objectiveoptimization problems. Different meaheuristics may have different features, such asdifferent intensification ability and diversification ability. Aim to combine the meritsof PSO and SA algorithms, this paper studies the hybridization of PSO and SAalgorithm for biclustering gene expression data, and studies the parallelization of themulti-objective SA algorithm based on PSO algorithm.In order to enhance the intensification ability of PSO algorithm, this paperpresents a hybrid multi-objective particle swarm optimization biclustering algorithmbased on simulated annealing ideas. The proposed algorithm integrates the idea ofsimulated annealing to PSO algorithm. Specifically, we use the metropolis criterion ofSA algorithm of decide whether to accept or reject particle’s speed and locationupdate. In this way, better solutions will be accepted always, and worse solutionswill have a chance to be accepted also. This strategy may balance the intensificationand diversification ability of the proposed algorithm more efficiently.Subsequently, this paper presents a hybrid multi-objective SA algorithm basedon the idea of PSO algorithm. Due to the random sampling strategy, traditional SAalgorithm is very slow in convergence speed. Recent studies on function optimizationproblem have showed that learning based sampling can improve the convergencespeed of SA algorithm significantly. In PSO algorithm, the speed and position update equation exhibit strong leaning ability. Aim to improve the sampling efficient of SAalgorithm, this paper uses the movement equation of PSO algorithm to producecandidate solution for SA algorithm. In this way, the proposed hybrid SA algorithmcan have an adaptive neighborhood, so it can search the solution space more finely.The intrinsic serialization of SA algorithm makes its parallel efficiency ofparallel implementation heavily dependent on whether the problem is parallelizable.The hybrid multi-objective SA algorithm for studying is a population-based algorithmwhich is intrinsic parallel. This paper studies the performance of the parallel hybridmulti-objective SA algorithm on processing gene expression data clustering analysis.Overall, the paper has conducted in-depth study on multi-objective optimizationbiclustering algorithm based on PSO and SA algorithm. Two different hybridstrategies are proposed to take advantage of the merits of PSO and SA algorithm, andwe implement the parallel version of hybrid multi-objective SA algorithm also. Allthose proposed algorithms are tested on two datasets of microarray gene expressiondata proposed, and the experiment results show that those algorithms are promising.