On-line sliding-window Levenberg-Marquardt methods for neural network models

Abstract

On-line learning algorithms are needed when the process to be modeled is time-varying or when it is impossible to obtain off-line data that covers the whole operating region. To minimize the problems of parameter shadowing and interference, sliding-based algorithms are used. It is shown that, by using a sliding window policy that enforces the novelty of data stored in the sliding window, and by using a procedure to prevent unnecessary parameter updates, the performance achieved is improved over a FIFO policy with fixed parameter updates. Important savings in computational effort are also obtained.