基于改进粒子群算法的过程神经网络训练

针对过程神经元网络由于模型参数较多,正交基展开后的BP算法不易收敛的问题,结合量子理论,提出一种改进的粒子群算法,并用于过程神经元网络的训练。算法中粒子采用双链结构,用量子位的概率幅对粒子位置编码,通过量子旋转门和量子非门完成粒子的更新与变异,可发挥量子粒子群的群体搜索能力和全局收敛性,有效克服BP算法计算复杂、容易陷入局部最小值等缺陷。以两组二维三角函数的模式分类问题为例,验证算法有效性。结果表明该方法不仅收敛速度快,而且寻优能力强。

Training of Process Neural Network Based on Modified Particle Swarms Algorithm

Aiming at the problem that it is difficult for BP algorithm to converge because of more parameters in training of process neural networks based on orthogonal basis expansion, a modified particle swarm algorithm is presented which combines the quantum theory and is to train the process neural network. Each particle has the double chain structure. The positions of particles are encoded by the probability amplitudes of quantum bits. The movements and mutations of particles are performed by quantum rotation gate and quantum non-gate. The algorithm has the abilities of population search for quantum swarm particle and global convergence, which overcomes the complex compute and easily plunges into local minimums about BP algorithm. Taking the pattern classification of two groups of two-dimensional trigonometric functions as an application, the simulation results show that the algorithm has not only fast convergence but also good optimization ability.