黄余

摘 要： 为克服传统BP神经网络在运算过程的不足，提出一种基于高维粒子群算法的神经网络优化方法。通过在高维PSO算法中引入随机变化的加速常数来获得最优权值，对BP神经网络进行优化和训练，再将优化好的高维BP神经网络运用到交通事件自动检测中，通过检测训练算法，并对训练后的数据进行分类测试，把分类测试的结果与传统BP神经网络和经典事件检测算法比较。结果显示，经过优化后的高维粒子群BP神经网络的检测率、算法性能均优于BP神经网络算法和经典算法，其中97，50个测试样本中仅有2个测试样本与应该达到的数值不一致，其他样本都满足测试要求，并且平均优化测试时间是传统BP神经网络检测时间的一半，因此，优化后的BP神经网络算法的性能十分优越。

关键词： 高维BP神经网络； 粒子群算法； 神经网络； 结构优化

中图分类号： TN711?34； TP311 文献标识码： A 文章编号： 1004?373X（2017）03?0157?03

Research on a neural network structure optimization method based on

high?dimensional particle swarm optimization

HUANG Yu1， 2

（1. College of Computer Science， Chongqing University， Chongqing 400044， China； 2. Yibin Vocational and Technical College， Yibin 644003， China）

Abstract： In order to eliminate the shortcomings of the traditional BP neural network in the operation process， a neural network optimization method based on the high?dimensional particle swarm optimization algorithm is proposed. The acceleration constant with random variation is introduced into the high?dimensional PSO algorithm to acquire the optimal weight to optimize and train the BP neural network. The optimized high?dimensional BP neural network is applied to the automatic detection of the traffic incident. The trained data is performed with class test with the detection and training algorithm， and its result is compared with those tested with the traditional BP neural network algorithm and classical event detection algorithm. The results show that the detection rate and performance of the algorithm optimized with high?dimensional particle swarm optimization BP neural network algorithm are better than those optimized with BP neural network algorithm and classical algorithm， the values of 2 test samples are different with the expected values of 97 and 50 test samples， the rest samples can meet the test requirement， and the average optimal testing time is half of the detection time of the traditional BP neural network. The optimized BP neural network algorithm has excellent performance.

Keywords： high?dimensional BP neural network； particle swarm optimization； neural network； structure optimization

0 引 言

目前关于高维多目标优化问题的理论和方法大都限于少数几种算法，如NSGA?II算法，粒子群优化（PSO）算法。粒子群优化算法是Kennedy等受到飞鸟集群活动的启发而提出的一类新兴的基于集群智能优化算法[1?5]。相比进化算法，PSO具有易于实现和收敛速度快等优势。近年来PSO在多目标优化领域的研究上取得了较大进展[6?9]，在神经网络训练、复杂函数优化、图像处理、工业系统优化等领域应用广泛[10]，PSO在特性的函数优化问题上更是发挥着重要作用，这些特性函数通常具备维数高、非线性、规模大、非凸和不可微等特点，一般计算方法运算困难。……