Particle Swarm Optimization Algorithm For Dynamic Optimization And Dynamic Clustering Problem

With the popularization of computer technology and network technology in real life, the optimization problems people facing become more and more complex, which raises more scientific research on the dynamic multi-objective optimization field. The optimization problem in real life is complicated, and the objectives are mostly dependent on the external environment, and the optimization method based on the theory of mathematics has been unable to deal the reality problem. Based on swarm intelligence, evolutionary algorithm can find the optimal solution or the optimal solution set by simulating the intelligent biological behavior.Data stream as a mainstream form of data, has the characteristics of large scale, fast producing speed and time label. It has been widely used in various fields of application. Mining the useful information in data stream has become a hot research field in data mining, thus a batch of data stream clustering algorithm has emerged.Based on these backgrounds, this paper presents a dynamic multi-objective optimization algorithm based on prediction and decomposition, and two automatic clustering algorithms based on particle swarm optimization algorithm for data stream. The work of this paper includes the following.1. A dynamic multiple populations particle swarm optimization algorithm based on decomposition and prediction(DP-DMPPSO) is proposed to deal with the dynamic multi-objective optimization problems. The algorithm adopts the multiple population co-evolution basic framework. Each sub-population optimizes one objective, and the population evolves independently. At each environment, the external archive will be output as the final result. Therefore, a mechanism based on objective space decomposition is proposed to update the archive, which makes the final solution set not only non-dominant, but also distribute evenly. In addition, a population based prediction mechanism is adopted to deal with the environmental changes. By predicting the initial population of the new environment, the convergence speed of the population will be faster. In order to test the performance of the algorithm, the DP-DMPPSO algorithm is used to optimize 6 sets of function, and the optimization results are compared with that of the other four dynamic optimization algorithms. There are 3 sets of dynamic test functions with complex Pareto solution set. The experimental results show that the prediction mechanism of DP-DMPPSO algorithm has predicted the Pareto front of complex functions, and the method of external warehouse update based on objective space decomposition also makes the solution set have better distribution.2. Two automatic clustering algorithms based on the particle swarm optimization for data stream clustering are proposed: the first one is based on manifold distance(MD-PSO) and the other one is based on density(Den-PSO). As the data stream is unknown, the number of categories cannot be pre-set, and so the automatic clustering can deal with this kind of data. MD-PSO algorithm adopts the variable length of the encoding method. According to the number of categories of each particle, the population is divided into sub-populations. Encoding represents the cluster center, and then cluster the dataset through a manifold distance based clustering method. According to the clustering results, fitness will be assigned for particles. After the evolution process, find the optimal solution of another objective function from the global optimal solutions of all sub populations. In Den-PSO algorithm, according to the density based clustering method, each particle also represents a clustering result which is used to evaluate the particles. After the evolution process, the optimal solution of another objective function is found from the history best position of all particles, and this optimal solution will be output as the final clustering result of the algorithm. These two algorithms are applied to the data stream clustering, and the clustering results show that the clustering results of both two algorithms are good. When these two algorithms meet the breakpoint, the Den-PSO algorithm can create a new core for the breakpoint data, which indicates Den-PSO is more automatic. These two algorithms are the founder of the evolutionary algorithm for data stream clustering.