Cognitive and social learning factors are two important parameters associated with the performance of particle swarm optimization significantly. Up to date, many selection strategies have been proposed aiming to improve either the performance or the population diversity. One of the most widely used improvements is the linear selection manner proposed by Ratnaweera in 2004. However, due to the complex nature of the optimisation problems, linear automation strategy may not work well in many cases. Since the large cognitive coefficient provides a large local search capability, whereas the small one employs a large global search capability, a new variant - predicted modified particle swarm optimization with time-varying accelerator coefficients, in which the social and cognitive learning factors are adjusted according to a predefined predicted velocity index. If the average velocity of one particle is superior to the index, its social and cognitive parameters will chose a convergent setting, and vice versa. Simulation results show the proposed variant is more effective and efficient than other three variants of particle swarm optimization when solving multi-modal high-dimensional numerical problems.
