Research On Group Search Optimization Algorithm And Applications

Many specific issues in the scientific researches and engineering designs could be summarized as parameter optimization problems. Recently, as the developing of swarm intelligent theory, the swarm intelligent algorithm which is inspired by the animals’behavior, becomes more and more popular. Moreover, the swarm intelligent algorithm under a big challenge caused by the super development of industry process, such as multi-modal, high-dimensional, on-linear and constrained problems. As a novel algorithm the Group Search Optimization (GSO) algorithm is gradually well known and adopted by its special swarm structure, also its outstanding performance on solving complex problems. However, the GSO is still not completely free from the problems such as entrapped by the local optima and premature convergence. This thesis aims at improving the GSO algorithm by various strategies to enhance its ability for solving complex problems. The main contributions of this thesis can be summarized as follows:(1) In order to enhance the global search ability of Group Search Optimization (GSO), a hybrid algorithm called Group Search Particle Swarm Optimization (GSPSO) based on GSO and the Particle Swarm Optimization (PSO) is proposed. The GSPSO combines the advantage of the PSO on searching in the global space and the good performance of the GSO on searching in a local area. In the GSPSO, the PSO model and the GSO model are used in turn. The PSO model is used to find a good local area, which usually contains or nearby the global optimization point. The swarm executes the GSO model search in the local area; by the way, the rangers are employed to revise the local area. For switching the two models, a mutual rescue method is also designed. Moreover, in order to increase the diversity of swarm, some members with weak fitness value are weeded out in each iteration.18benchmark functions are used to evaluate the performance of the novel algorithm. The results show that the GSPSO has better convergence accuracy for most problems compared to other four algorithms.(2) In GSO the scroungers move to global best member directly causing to premature convergence when solving global optimization problems. Thus in order to increase the diversity of scroungers’ behavior an improved group search optimization (ISWGSO) is proposed by building a neighborhood by small world network. In ISWGSO, each scrounger builds up its own neighborhood according to the small world scheme, and evolves with the effects of both global best member and local best member in its neighborhood. The scale of neighborhood increases during the whole calculation; moreover a dynamic probability scheme is designed for adding neighbors, in which the distance between each member is a key factor. Moreover, factorial design (FD) approach is used to select parameters of ISWGSO for different problems. The result of numerical examples on eleven benchmark functions shows that ISWGSO can obtain a satisfied performance on multi-modal problems. Finally, ISWGSO is applied to train the parameters of neural networks to build a soft sensor model for inferring the outlet ammonia concentration in fertilizer plant.(3) Compare to the low-dimensional problems, the multi modal problems with high dimension are more pregnant. Facing this kind of problems, although GSO shows better performance than other algorithms, it is still not completely free from the problem of premature convergence. Therefore, for solving the high-dimensional problem, a GSO using a neighborhood (GSO-NH) is proposed in this paper. The topology of neighborhood is inspired from the animal collective behavior and be utilized to each group member. Based on the neighborhood, members abandon the angle evolution strategy, and novel strategies for the three roles are adopted. The producer explore a better position by the help other outstanding neighbors. To enhance scroungers’ ability on exploration, the local best members in neighborhoods are used to influence scroungers by coalition with the global best member. Moreover, in order to balance the exploration and exploitation on different problems, two weights are employed to balance the trust on the local and global best member. To enhance the search behavior of rangers, a mechanism of mutation with a controlling probability is introduced. In addition, for this mechanism, the bounds of rangers’ neighborhoods are used to generate dynamic searching steps. Over a suit of benchmark functions, GSO-NH is shown to be more effective and robust than or at least comparable to other five algorithms, especially on high-dimensional multimodal problems. Finally, the GSO-NH is used to train a neural network with337parameters, and the result shows that GSO-NH is applicable for real-world high-dimensional problems.(4) For the constrained problems, an improved algorithm named GSO-TR is also proposed. Inspired by the tandem running behavior of ant colony, the infeasible individuals are divided into sharers and explorers. In order to increase the diversity of swarm in the feasible area, the shares adopt a neighborhood position copy strategy to move in the feasible area directly. The explorers excavate the information near the edge of feasible area, by approaching to the feasible area step by step. Moreover, the toleration to the infeasible individuals is retrenched by increasing the proportion of sharers. Tested by some classical constrained problems, the GSO-TR shows to be a potential method to solve the constrained problems. Moreover, the GSO-TR is also applied to solve a real-world problem for the simplified alkylation process. The promising results on this problems show that the GSO-TR is applicable for real-world problem solving.