Abstract: An improved method for RBF neural network parameters optimization is proposed. The number of nodes in the hidden layer is determined by using RAN（Resource Allocating Network）, meanwhile strategy of pruning is introduced to remove those hidden units which make insignificant contribution to overall network output. Central position, width and weight of the neural network are optimized by the improved PSO（Particle Swarm Optimization） algorithm, so as to obtain the appropriate structure and control parameters. The new algorithm is used to predict the model of CSTR, and the result indicates that RBF neural network optimized by this algorithm has a smaller structure and high generalization ability.

Key words: radial basis function neural network, resource allocating network, strategy of pruning, particle swarm optimization

摘要： 提出了一种改进的RBF神经网络参数优化算法。通过资源分配网络算法确定隐含层节点个数，引入剪枝策略删除对网络贡献不大的节点，用改进的粒子群算法对RBF网络的中心、宽度、权值进行优化，使RBF网络不仅可以得到合适的结构，同时也可以得到合适的控制参数。将此算法用于连续搅拌釜反应器模型的预测，结果表明，此算法优化后的RBF网络结构小，并且具有较高的泛化能力。

关键词: 径向基神经网络, 资源分配网络, 剪枝策略, 粒子群优化