There are many adaptive signal processing applications which attract a lot of researchers in recent years. For example: the echo cancellation, noise control and etc. Because of the simplicity, LMS is very popular in the real applications. One of the important research trend of LMS is to study the variable length of the filter. In the thesis, we propose a variable tap-length adaptive filter algorithm. It uses on the gradient of error signal to tune filter to suitable order. It is an improved GD algorithm. It has better performance than original GD algorithm in any aspect.