Improvements And Applications Of Particle Swarm Optimization Algorithm

Particle swarm optimization (PSO) is one of the most important swarm intelligence techniques. PSO is based on simulations of social behaviors such as animals herding, fish schooling, and birds flocking where the swarms search for food in a collaborative manner. Since PSO is easy to implement and has higher search efficiency, it develops greatly in recent years. Many successful applications on solving real-world optimization problems by PSO can be found in recent literatures. This dissertation is focused on the research of parameters adjustment in PSO algorithms, and several novel strategies on adaptive parameters adjustment in PSO are proposed by this dissertation. Experimental results on Benchmark functions demonstrate the effectiveness of the improved PSO algorithms in solving complex optimization problems. Afterwards, several PSO algorithms are applied to the real-world engineering design constrained optimization problems, multilevel thresholding for image segmentation and change detection for SAR images. The main innovative points of this dissertation can be summarized as follows:1. Two improved PSO algorithms with adaptive parameters are proposed in this dissertation. The first algorithm is a variant of the PSO that we call the adaptive particle swarm optimization with dynamic population (DP-APSO), which adopts a novel dynamic population (DP) strategy whereby the population size of swarm can vary with the evolutionary process. The DP strategy enables the population size to increase when the swarm converges and decrease when the swarm disperses. Experiments were conducted on twelve benchmark functions including both unimodal and multimodal problems. The results demonstrate perfect performance of the DP-APSO in solving complex multimodal problems when compared with six other variants of the PSO. The second algorithm is called the dynamic population PSO with consecutive generations strategy and uniform mutation operator (DPPSO-CG-UM). The population size of DPPSO-CG-UM increases or decreases by one at run time according to the judgment of the CG strategy, and the new particle inserted into the swarm will be produced by the UM operator. Experiments were conducted on ten benchmark functions including multimodal and unimodal problems. The experimental results demonstrate competitive performance of the DPPSO-CG-UM on solving complex multimodal problems especially when compared with several other variants of the PSO.2. Since many real-world optimization problems involve constraints, constrained optimization are used widely in engineering applications. A new approach for solving constrained optimization problems is presented, which is named as the simple multi-population evolutionary algorithm using PSO strategy (PSO-SMEA). In the PSO-SMEA, a constrained optimization problem is converted into several unconstrained optimization problems by parallel evolutions of two populations, which are dynamically formed. The PSO-SMEA is applied to solving four well-studied engineering design constrained optimization problems (Welded Beam Design, Spring Design, Speed Reducer Design and Three-Bar Truss Design), and the results show that the PSO-SMEA is competitive when compared with two other constrained optimization algorithms.3. Since the conventional multilevel thresholding approaches exhaustively search the optimal thresholds to optimize the objective functions, they are computational expensive. The modified particle swarm optimization (MPSO) algorithm is proposed to overcome this drawback. The MPSO employs two new strategies to improve the performance of original PSO, which are named the adaptive inertia (AI) and adaptive population (AP) respectively. With the help of AI strategy, the inertia weight is variable with the searching state, which helps the algorithm to increase search efficiency and convergence speed. Moreover, with the help of AP strategy, the population size of MPSO is also variable with the searching state, which mainly helps the algorithm to keep diversity and jump out of the local optima. The MPSO algorithm is used to find the optimal thresholds by maximizing the Otsu’s objective function, and its performance has been validated on twelve standard test images. The experimental results of30independent runs illustrate the better solution quality of MPSO when compared with the global particle swarm optimization (GPSO) and standard genetic algorithm (SGA).4. A new fuzzy clustering algorithm using multilevel thresholding is proposed to reduce the computational complexity of fuzzy local information c-means (FLICM) algorithm for solving the clustering problem on the difference image of change detection for SAR images. First, the pixels in the difference image are classified into the "changed" pixels,"unchanged" pixels and unknown status pixels by the multilevel thresholding procedure based on the PSO. Then the unknown status pixels are clustered by the FLICM. If the neighboring pixels in the FLICM are not the unknown status pixels, their degrees of membership are set to1or0according to the results of multilevel thresholding before. The proposed algorithm improves the precision in the change detection for SAR images with the low computational complexity. It can be concluded by the experimental results that the proposed algorithm has the better performance than fuzzy c-means (FCM) and FLICM on the change detection for SAR images and its run time is about70%less than that of FLICM.5. Change detection for SAR images can be transformed into the clustering for the difference image of SAR images. Since SAR images have speckle noise, a new adaptive particle swarm clustering algorithm using neighborhood information is proposed for improving the clustering results. The degrees of membership of the neighbors around each central pixel are introduced into the new objective function based on the FCM clustering algorithm, and the centers of clusters are optimized by the global searching of adaptive particle swarm. By the self-study operator of the proposed algorithm, the degree of membership of each pixel can be revised based on the degrees of membership of all the neighboring pixels. It can be concluded by the experimental results that the proposed algorithm is less sensitive to noise than FCM and quantum-inspired immune clonal clustering algorithm by reason of the using of neighborhood information, and is better than FLICM on image detail preservation and run time.