LMS algorithm with an adaptive neural network cost function

We propose a new LMS algorithm with an adaptive neural network cost function (ANNCFLMS) for application to unknown channel estimation or system identification. The algorithm employs the weighted average of a neural network with two input signals--the squared errors at adjacent time intervals--to modify the cost function and update the respective weight according to a gradient descent algorithm designed to track the minimum mean squared error (MSE). For fast convergence, the step-size updates recursively until the modified cost function attains its minimum value. Simulation results demonstrate that the proposed algorithm converges faster and is especially robust in low-SNR or colored input environments.