Suitable For The Study Of Particle Swarm Optimization, Stochastic Optimization Problems

There is much objective or man-made randomness in the fields of management science, information science, systems science, industrial engineering and so on, where a large number of stochastic optimization problems exist accordingly. Although evolutionary computation (e.g. genetic algorithm, evolutionary strategy, particle swarm optimization etc.) has been showed to be powerful global optimizers for stochastic optimization problems, it has the disadvantage of higher cost which limits its applications because all individuals'fitness is estimated by random simulation. It is suggested to predict the fitness of some individuals, taking the information of the fitness that has been calculated into account. This method can reduce the times of random simulation, and accordingly it reduces the computing time in total.Firstly, generalized regression neural network (GRNN) is used as a fitness prediction model and an intelligent algorithm combined GRNN with particle swarm optimization is presented. In this intelligent algorithm, according to the analysis of the GRNN, the mechanism which combines GRNN with particle swarm optimization, prediction strategy and prediction model training are proposed, by which we can decide whether the individual's fitness is predicted by the GRNN or estimated by random simulation and when to update the GRNN. Results of simulation show that the intelligent algorithm reduces the computational cost greatly in the premise of performance guaranteed. Next, Kriging model is used as a fitness prediction model and a new algorithm which combines Kriging model with particle swarm optimization is proposed. In this algorithm, according to the mechanism combined Kriging model with particle swarm optimization and prediction strategy, some of the individuals'fitness is predicted and the rest is estimated by random simulation. Results of simulation show the correctness and effectiveness of this new algorithm.