Performance Analysis And Optimization For Particle Swarm Optimization

As a population-based stochastic optimization algorithm, particle swarm optimization (PSO) simulates the social behavior among animal society such as bird flocking and fish schooling. Different from other evolutionary stochastic optimization algorithm, it employs not only the position information, but also the velocity information to control the particles'trajectories. Due to the easy implementation and the fast convergent speed, it has been successfully applied into many areas. In this thesis, the reason affecting the performance are analyzed, and several improvements, such as structure optimization, parameter selection, hybrid algorithm and estimation of fitness, are designed to make PSO more effective.In standard PSO, the step is a constant when the differential model is translated into difference model, this limitation affects the performance greatly due to a large error between these two models. Therefore, this thesis proposes the differential evolutionary PSO model with an additional parameter– step, and the stochastic selection strategy is established with absolute stability theory. Because the prey time is very short compared with the whole searching food procedure, this selection strategy provides a more chance to observe the prey procedure, and is more fit for the biological background. Based on different numerical method such as Euler method, modified Euler method and Runge-Kutta method, three differential evolutionary PSO algorithms are designed, simulation results show they are effective and efficiency especially for Runge-Kutta method.With the difference model of PSO, this thesis analyzes the structure with Z-translation, and finds the standard PSO can be viewed as a feedback system with two inputs and one output. Then, the controller is incorporated into the standard PSO to construct a new PSO model with controller. Furthermore, with different controllers such as integral controller and PID controller, the corresponding variants are designed, and the parameters'selection principles are obtained through stability analysis and support set theory. Simulation results show this type of PSO can improve the global exploration capability significantly.As an important parameter, the affection of velocity threshold is analyzed from the viewpoint of convergence and calculation efficiency. Then, two effective velocity threshold adjustment strategies are designed: the stochastic strategy and individual strategy. The first one adjusts the ratio between the global exploration capability and the local exploitation capability randomly, whereas the second one investigates the velocity threshold setting with individual-charactered behavior inspired from biological background. Simulation results show both of them are effective when solving the control problem of chaotic system.To design an effective hybrid algorithm combined with PSO and mutation operator, the relationship between the randomness of cognitive component and the local exploitation capability is analyzed with linear control theory. In order to improve the calculation efficiency, the randomness is eliminated from the cognitive component to reduce the global search capability, as well as enhance the local search capability. The application result of non-stable linear system approximation is superior than other previous reports.For applications those need a large number of fitness calculation, this thesis proposes two estimation strategies. The first one makes a random estimation with weighted mean of fitness, while the second one estimates the fitness based on the reliability value. Both of them are effective and have been successfully applied to uncertain programming.