Removing erroneous history stack elements in concurrent learning

This paper is concerned with erroneous history stack elements in concurrent learning. Concurrent learning-based update laws make concurrent use of current measurements and recorded data. This replaces persistence of excitation by a less restrictive linear independence of the recorded data. However, erroneous or outdated data prevents convergence to the true parameters. We present insights into the convergence properties of concurrent learning and propose a routine to recognize and remove erroneous data online. We characterize erroneous data based on its inconsistency with the current measurement-based update. We numerically validate that the proposed routine restores the tracking ability and improves the convergence properties of concurrent learning.