| From the 1980s, intelligent Optimization algorithm such as neural network, GA, chaos has been developed through the simulation of nature and social process and it presents a new approach for optimization methods. Particle swarm optimization is a population-based, self-adaptive search optimization technique. As a kind of intelligent algorithm, it can be used to solve various optimization problems and shows great potential in practice. Now, it has been widely applied in many other areas, such as artificial neural network and fuzzy system control.This thesis presents an analysis of the convergence behavior of the PSO, and dig into the premature convergence problem. In order to overcome this problem, two modified PSO -SDPSO and VMPSO are proposed. SDPSO, motivated by nature and physics principles, can maintain a high level of diversity so as to obtain the ability of continuously searching. The results of experimental simulations show that SDPSO can not only overcome premature convergence problem effectively, but also increase the probability for the swarm to jump out of the local minimum. VMPSO adjusts the velocity of the particle by dimension, which can facilitate high-dimension space searching. The experimental results show that for 30-D Griewank and Rastrigin functions the algorithm can even find the global optimum - zero. For all testing functions in this experiment, VMPSO outperform PSO a lotThe two modified algorithms are applied to the task of training two-layer perceptron for three benchmark classification problems and compared with other learning algorithms such as BP, GD and GA. The experimental results illustrate the efficiency. The thesis also presents a solution of creating cognitive states classifier using PSO-based perceptron network. A classifier is being established and trained according to the solution to distinguish between two Chinese cognitive states: Chinese Character and Pinyin Reading. This solution can also be applied to solve various classification problems.
|
PDF Full Text Request