Abstract: Aiming at the problem that the modeling accuracy of neural network power amplifier is not high and easy to fall into local extremum, a new improved parallel particle swarm optimization algorithm （Improved Parallel Particle Swarm Optimization, IPPSO） is proposed. The adaptive mutation operation is introduced into the improved algorithm based on the parallel particle swarm algorithm, which avoids falling into local optimum. Meanwhile, the global optimal position of the population is added to the speed of the particles, and it adjusts learning factor adaptively and linear decreasing inertia weight to speed up the convergence of particles. Finally, the improved algorithm is used to optimize the parameters of RBF neural network, and the network is used to model the nonlinear power amplifier. Compared with the standard particle swarm algorithm, the root mean square error of this method is improved by 19.08%, which verifies the feasibility of the algorithm and improves the accuracy of the neural network power amplifier modeling effectively.

Key words: parallel particle swarm optimization, adaptive mutation operation, Radial Basis Function（RBF） neural network, average fitness, power amplifier modeling

摘要： 针对已有神经网络功放建模的建模精度不高，易陷入局部极值等问题，提出一种新的改进并行粒子群算法（Improved Parallel Particle Swarm Optimization，IPPSO）。该算法在并行粒子群算法的基础上引入自适应变异操作，防止陷入局部最优；在微粒的速度项中加入整体微粒群的全局最优位置，动态调节学习因子与线性递减惯性权重，加快微粒收敛。将该改进算法用于优化RBF神经网络参数，并用优化的网络对非线性功放进行建模仿真。结果表明，该算法能有效减小建模误差，且均方根误差提高19.08%，进一步提高了神经网络功放建模精度。

关键词: 并行粒子群算法, 自适应变异操作, 径向基函数（RBF）神经网络, 平均适应度, 功放建模