一种新型模拟退火神经网络及其应用

Neural Network with New Simulated Annealing and Its Applications

针对训练神经网络权值的BP算法容易陷于局部最小值点的问题,提出了带自适应冷却进度表的模拟退火算法与Powell算法构成新型混合算法,用该算法训练网络的权值。冷却进度表中主要参数是模拟退火算法的控制参数T的初值T0和T的衰减函数。把整个迭代过程划分为若干阶段,在每个阶段结束时,依据网络模型误差自适应地修正下阶段的T0(回火温度)、T的衰减函数中的参数和迭代步长初值。仿真结果表明,上述混合算法具有很强的全局和局部搜索能力,其性能优于BP算法;该算法在油田系统建模问题中的成功应用也表明了该方案的有效性。

In order to overcome the problem of BP algorithm falling into local minimum, a hybrid algorithm is proposed to train weights of neural netwoks. The algorithm consists of Powell algorithm and simulated annealing algorithm with adaptive cooling schedule. The main parameters in the cooling schedule are initial value To and the attenuation function of control parameter T in SAA. The whole iteration is divided into several stages. When each stage ends, according to NN model errors, it is adaptively modified that next stage＇s T0 （tempering temperature）, and the parameter of attenuation function of T, and initial value of iterative step-length. Simulation results show that the above hybrid algorithm has strongly global and local search property, and advantage over one of BP algorithm. The algorithm is applied to oil field system modeling, which proves that the proposed scheme is useful and effective.