On the basis of data mining,a new method is developed for identifying fuzzy model,updating its parameters and determining optimal division of output space simultaneously by means of fuzzy sets theory and the improved gradient descent method.

Levenberg Marquardt (LM) algorithm is a combination of neural network, Gauss Newton method and gradient descent method, so it has the abilities of self learning, fast convergence and global search.

Two ways are used to design the network,the one is the direct energy descent method,and the other is the gradient descent method.

Comparison between GA and gradient descent algorithm in parameter optimization of UPFC fuzzy damping controller;

A few or all of the parameters of the controller are adjusted by using the gradient descent algorithm to minimize the output error.

A robust speech recognition system based on discriminative learning of environmental features is proposed for recognition of environmental features in high noise background, and a gradient descent algorithm was adopted for parameters optimization.

Aimed at the feature of weak local searching ability of the genetic algorithm,let it be combined with the gradient descending method so as to raise its partial searching ability,and a test has been carried out.

Then,the gradient descending method is applied to optimize the weights by iterative arithmetic to minimize the errors of objective value.

It needs to work out parameters of the base function in the process of finding bias field,but conventional methods such as gradient-descent method often find local best.

The authors estimate the bias field by Legendre polynomials to find the parameters with minimum entropy,conventional ways such as gradient-descent method often find local best,to find global best,the authors present genetics algorithm to find best parameters to estimate the bias field,but it can not always find global best neither.

During the procedure of the iterations,the gradient descent operator is designed to accelerate the convergent speed and the mutation operator to improve the diversity of particle swarm.

Therefore,this article puts forward to new dynamic neural networks with integrator and adjustable feedback coefficient and gets its learning algorithms by gradient descent.