Fixed-budget kernel least mean squares

In this paper a new version of kernel normalized least mean squares algorithm is proposed, allowing for maintaining a fixed memory budget. Constant growth of the support vectors dictionary, inherent to online kernel methods is constrained using pruning criterion. After admitting a new input vector to the dictionary, the least important entry is found and removed. This allows for competition among the input vectors and dynamic adaptation of the dictionary content. In the task of time-varying system identification this method has the clear advantage over a coherence criterion which does not modify vectors already present in the dictionary. Experimental results confirm that algorithm performs well in time-varying environment and it can maintain a performance comparable to the state-of-the-art algorithms, using smaller number of support vectors, with linear complexity.