Enhanced Particle Swarm Optimization Based on Reference Direction and Inverse Model for Optimization Problems

While particle swarm optimization (PSO) shows good performance for many optimization problems, the weakness in premature convergence and easy trapping into local optimum, due to the ignorance of the diversity information, has been gradually recognized. To improve the optimization performance of PSO, an enhanced PSO based on reference direction and inverse model is proposed, RDIM-PSO for short reference. In RDIM-PSO, the reference particles which are used as reference directions are selected by non-dominated sorting method according to the fitness and diversity contribution of the population. Dynamic neighborhood strategy is introduced to divide the population into several sub-swarms based on the reference directions. For each sub-swarm, the particles focus on exploitation under the guidance of local best particle with a good guarantee of population diversity. Moreover, Gaussian process-based inverse model is introduced to generate equilibrium particles by sampling the objective space to further achieve a good balance between exploration and exploitation. Experimental results on CEC2014 test problems show that RDIM-PSO has overall better performance compared with other well-known optimization algorithms. Finally, the proposed RDIM-PSO is also applied to artificial neural networks and the promising results on the chaotic time series prediction show the effectiveness of RDIM-PSO.