Abstract: To improve the method to interpose structure and parameters of neural network, a novel Radial Basis Function（RBF） neural network method based on Improved Particle Swarm Optimization（IMPSO） is proposed. The convergence speed of this algorithm and the capacity of searching global optimum value are increased through adjusting inertia weight factor dynamically. The experiments show that the neural network based on IMPSO algorithm is superior to self-organizing center selected algorithm and standard PSO algorithm in the capacity of function approximation, and enhances the generalization and the optimized effect of the network. The capacity of solving nonlinear problems of this algorithm is enhanced effectively.

Key words: Particle Swarm Optimization（PSO）, Radial Basis Function Neural Network（RBFNN）, inertia weight factor

摘要： 为了改进神经网络模型结构和参数的设置方法，提出了一种改进的粒子群优化径向基函数（RBF）神经网络的方法。该方法通过动态调整粒子群算法中的惯性权重因子，提高了算法的收敛速度和搜索全局最优值的能力。实验结果表明：基于改进的PSO算法训练的神经网络在函数逼近性能上优于自组织选取中心算法与标准PSO算法，提高了网络泛化能力和优化效果，有效地增强了网络对非线性问题的处理能力。

关键词: 粒子群算法, 径向基神经网络, 惯性权重因子