A Reliable and Efficient Hybrid PSO for Parameters Optimization of LS-SVM in Production Rate Prediction

Least square support vector machine (LS-SVM) has become an effective tool in nonlinear function estimation. But it is a hard optimization problem to determine kernel parameters for LS-SVM owing to its implicit form and numerous local optima. A reliable and efficient hybrid PSO algorithm named self-adaptive lattice PSO with chaotic operator (short for cPSO) is proposed, which can obtain the same results in different runs. Therefore, once the appropriate algorithm parameters are determined for some practical problem, it will always achieve satisfactory result in every run. In cPSO algorithm, a new chaotic operator is designed to replace random operators for population initialization and coefficients setting, which is a more efficient pseudo-random search strategy. To address the balance problem of standard PSO between exploration and exploitation, a bi-population strategy is adopted, in which one population search by Clerc's Constriction PSO with excellent convergence ability, and the other population performs self-adaptive lattice search with outstanding global exploration ability. The proposed method is used to establish the LS-SVM based predictive model between concentrate yield and the technical indices of all production procedures in the mineral dressing process. The practical results show that the proposed cPSO performs better than other optimization methods regarding both convergence accuracy and reliability.