Conclusions

We presented a new fixed-budget kernel recursive least-squares algorithm for online identification of nonlinear systems. To maintain its memory size, it combines a growing memory with a discarding criterion previously proposed for LS-SVM. We also presented an efficient updating method for pruning an arbitrary point from the dictionary, and a label update procedure to provide tracking capability.

The proposed method represents a significant improvement over the previously proposed SW-KRLS algorithm, and given similar memory requirements it also outperforms ALD-KRLS. Moreover, it is capable of tracking changes of a time-varying nonlinear mapping. Future research topics include the study of more sophisticated pruning criteria, and a comparison with other nonlinear trackers.