AbstractMotivated by the advances in computer technology and the fact that the batch/block least‐squares (LS) produces more accurate parameter estimates than its recursive counterparts, several important issues associated with the block LS have been re‐examined in the framework of on‐line identification of systems with abrupt/gradual change parameters in this paper. It is no surprise that the standard block LS performs unsatisfactorily in such a situation. To overcome this deficiency, a novel variable‐length sliding window‐based LS algorithm, known as variable‐length sliding window blockwise least squares, is developed. The algorithm consists of a change detection scheme and a data window with adjustable length. The window length adjustment is triggered by the change detection scheme. Whenever a change in system parameters is detected, the window is shortened to discount ‘old’ data and place more weight on the latest measurements. Several strategies for window length adjustment have been considered. The performance of the proposed algorithm has been evaluated through numerical studies. In comparison with the recursive least squares (RLS) with forgetting factors, superior results have been obtained consistently for the proposed algorithm. Robustness analysis of the algorithm to measurement noise have also been carried out. The significance of the work reported herein is that this algorithm offers a viable alternative to traditional RLS for on‐line parameter estimation by trading off the computational complexity of block LS for improved performance over RLS, because the computational complexity becomes less and less an issue with the rapid advance in computer technologies. Copyright © 2004 John Wiley & Sons, Ltd.