Research And Application Of Vertical Parameters Multi-swarm Particle Swarm Optimization Algorithm

Multi-modal Function Optimization problem is a complex optimization problem. Using traditional optimization methods to solve such problems encounters many difficulties, and swarm intelligence algorithms provide a new way of thinking to solve such complex optimization problems.PSO, as a swarm intelligence algorithm, has distinctive characteristics of swarm intelligence algorithm. In the multi-modal function optimization problems, multi-swarm particle swarm optimization algorithms have shown their simplicity and superiority. At present, in the existing research results, the division of swarms was on basis of fitness and neighborhood, and the number of swarms should be determined in advance. There are few studies on information exchanging between particles and between swarms and on effective methods of measuring swarm diversity.In order to solve the above-mentioned problems, this dissertation researches on learning characteristics of swarms, and proposes an algorithm named Vertical Parameter Multi-Swarm Particle Swarm Optimization algorithm (VPMSPSO). The main research contents include the following:Firstly, inspired by the individuals gradually forming a small groups based on learning from their neighbors and on mutual convergent, the dissertation proposes a vertical particle swarm optimization (VPMSPSO), and then researches on the convergence of VPMSPSO and proves it using random search theory. The results show that average convergence rate of VPMSPSO is slower than that of the standard particle swarm algorithm, but it can converge to the various peaks of multimodal function.Secondly, under the premise of that the swarm optimal value is a random sequence, the dissertation establishes nonlinear discrete-system model of VPMSPSO, and analyses swarm-merging influence on particle behavior and the system stability conditions, and then obtains selection range of the approximation degree threshold.Thirdly, by analyzing the reason of lost peak and solitary particle swarm in the evolutionary process, this dissertation proposes the concept of swarm relative difference degree and of swarm relative entropy, that measure swarm diversity, and improves the algorithm. The results show that the improved algorithm can better describe dynamic evolution of swarm, and effectively improve the diversity of swarm to avoid the phenomenon of lost peak and solitary peak. Fourthly, based on the dynamic clustering features of VPMSPSO, the dissertation designs a self-adaptive clustering algorithm, and applies to Gaussian RBF net construction, and designs a VPMSPSO classifier using natural number coding based on VPMSPSO.Finally, VPMSPSO algorithm is applied to the training neural network ensemble. For parallel search is a major feature of VPMSPSO algorithm, the training can not only obtain the network parameters but also do the number of ensemble network. According to the characteristics of water inflow data and gas emission data of the coal mine, the dissertation builds two kinds of water inflow forecasting models, including neural network ensemble model based on VPMSPSO algorithm training and RBF net model based on VPMSPSO self-adaptive clustering, and builds a gas emission concentration rules extraction model based the VPMSPSO classifier. Consequently VPMSPSO algorithm is applied to coal mining enterprise decision support early warning systems.