Research On Pool - Based Distributed Particle Swarm Optimization

As a stochastic optimization algorithm based on population, Particle Swarm Optimization algorithms needs a large number of time to calculate its fitness in the iterative process. This action wastes a lot of time and has a bad influence on efficiency of Particle Swarm Optimization algorithm. In addition, PSO converges slowly and it is prone to premature convergence.In order to solve these problems, this paper presents a kind of distributed PSO which uses a great number of computers calculating together to shorten calculation time and increase efficiency. In addition, asynchronous model of pool-based PSO is able to overcome defect of premature convergence,wide the range of searching, and improve the speed of convergence.At first,this paper makes a conclusion about the study of PSO these years. Then we proposes a pool-based distributed PSO(PBPSO) in order to solve problems in the present study. In PBPSO,several nodes or islands share a pool, and different clients deal with the data in pool together,until they finish all the task. PBPSO is achieved by several net-connected computers with evolutionary mechanism data sharing mechanism and particle locking mechanism. Finally, this paper analyzes the performance of the distributed system through the results getting from simulation.In the part of simulation, this paper constructs a distributed lab environment in the view of parallel computation. In fact, it is a distributed system in which several computers could work together. To test the distributed system,this paper chooses 4 test functions. In the end, this paper compares results with previous PSO. The results show that after several independent experiments our distributed system is better than previous PSO. Under the same conditions, the times of convergence become more, the average fitness has greatly improved, and the operation of algorithm reduces a lot of time-consuming. Multiple computing nodes working together shorten the calculation time, data pool and asynchronous model improve the performance of global convergence.