基于改进共轭梯度理论神经网络优化算法研究

文章阐述了一种改进共轭梯度理论神经网络优化算法。该方法是在传统共轭梯度算法(CG)基础上引入对输出权值进行优化的理念,故称其为输出权值优化共轭梯度算法(OWO-CG)。这种算法在进行学习时,首先根据误差函数利用共轭梯度法计算收敛因子,并修改输入层和隐含层的权值因子;接着,计算隐含层输出函数,利用相关输出权值优化理论构建并求解线性方程组得到输出层的权值因子;最后,计算误差函数,利用该算法不停地修正神经网络回路输出值与期望输出值之间的差值,直到满足精度要求为止。仿真验证结果表明,与传统共轭梯度算法相比,这种算法的学习过程更加迅速和准确。

The algorithm optimization based on improved conjugate gradient theoryof neural network

Structure based on neural network model of three layer network connection as basis, this paper proposes a new optimization algorithm, this algorithm for the conjugate gradient algorithm to improve the traditional, it is the output weight optimization algorithm based on conjugate gradient algorithm (OWO) and (CG) combining theory is proposed, so we call as the output weight optimization conjugate gradient algorithm (OWO-CG). The new algorithm combines two kinds of algorithm in a body, the whole learning process more quickly and accurately. Algorithm to learn every time the process can be divided into three steps: firstly, according to the error function, convergence factor by using conjugate gradient method, only the changes in input layer and hidden layer weights factor. Then, the output function layer unit calculation, using the output weight optimization output weighting factor theory and solving linear equations in order to get. Finally, the calculation error correcting output function, neural network circuit using the algorithm keeps the value of the difference between the output values and expectations, until meet the accuracy requirements. The experimental results indicate that, compared with the conjugate gradient algorithm and output weight optimization method, this algorithm greatly improves the training speed.