Particle swarm optimization (PSO) is a typical intelligence optimization algorithm, Because of the simple and easy implementation, it has great potential in practice and been widely applied in many areas, such as object function optimization, neural networks, combination optimization, image processing etc. However, PSO has some drawbacks such as premature, weak local exploration ability and low accuracy in optimal solution. In order to overcome its shortcomings, we proposed two kinds of PSO and applied in constraint optimization problems and BP neural networks respectively.In Chapter 1, we first summarize traditional numerical methods for solving constraint optimization problems such as penalty method, barrier function method, feasible direction method, multiplier method and so on. A comparision is then made between the traditional algorithm and modern algorithm in solving constraint optimization problems. Finally, the basic knowledge of BP neural networks and BP algorithm is introduced.In Chapter 2, we introduce PSO in the aspect of model, choice of parameters, research progress and so on. In particular, we show the practical application examples of PSO, which provide reference for the further application of PSO.Since it's inevitable to produce infeasible points when PSO is used to solve constraint optimization problems in Chapter 3, we first proposed a new strategy to deal with infeasible points. An hybrid particle swarm optimization (HPSO) is then proposed by treating differential evolution algorithm as the sub-algorithm of PSO. Finally, the numerical experiment indicated that HPSO could solve constraint optimization problem efficiently.In Chapter 4. we proposed a simple multiple population particle swarm optimization (SMPSO) whose every population individually explored in space. Furthermore, to avoid premature, we make the local optimal perturbation operation on the current optimization points by means of deviation from the mean of particle's fitness. This improves the global convergence of PSO. As an applicition, SMPSO is used to train the BP neural networks, and obtained satisfying pattern recognition results. |