Abstract: This paper proposes a new learning algorithm for process neural networks (PNNs) based on particle swarm optimization (PSO), called PSO-LM. After the orthogonal basis function expansion to the input functions and the weight functions of the PNN, the structure parameters and other parameters in the PNN will be formed as a particle, and globally optimized by PSO. This algorithm does not need any gradient calculations or the manual control of the network?s structure. The global learning capability and the convergence capability of the PNN can be guaranteed by the capabilities of PSO, so the PSO-LM can better develop and improve the approximation capability of the PNN. According to two practical prediction applications, PSO-LM can outperform the existing basis function expansion based learning algorithm ( BFE-LM) for PNNs, and the classic back propagation neural networks (BPNNs) on predictive accuracy.

Key words: process neural networks, learning algorithm, particle swarm optimization, basis function expansion

摘要： 基于粒子群优化为过程神经元网络提出了一种新的学习算法。新算法在对网络输入函数和连接权函数进行正交基函数展开后, 将网络中的结构参数和其他参数整合成一个粒子, 再用粒子群优化算法进行全局优化。新算法不依赖于函数梯度信息, 不需要手动调节网络结构。粒子群优化具有良好的全局优化性能和收敛性能, 保证了过程神经元网络的全局学习能力和新学习算法的收敛能力, 更好地发挥过程神经网络的逼近性能。两个实际预测问题的实验结果表明, 基于粒子群优化的学习算法比现有的基于梯度的基函数展开方法 以及误差反传神经网络模型具有更好的预测精度。

关键词: 过程神经元网络, 学习算法, 粒子群优化, 基函数展开