Based on building up a model of echo canceller, the convergence of gradient-type stochastic adjustment algorithm of an adaptive filter under the mean-squared error criterion is discussed.

All of the six ratios are statistically significant converge to the mean of industry, but the speeds of convergence are different from each other for the existence of adjustment cost.

The algorithm can remove redundant link even nodes on the network, through the network learning step dynamic adjustment to avoid convergence speed of.