In this thesis, background theory about the online kernel-based algorithms
and their use for online learning is presented. The analysis of the state-ofthe-
art methods highlights an important drawback in many kernel online
learning algorithms. This is the large memory storage needed due to the
amount of support vectors generated. We study the SCA approach for reducing
support vectors in the batch learning case and propose its adaptation
to the online scenario.
POLSCA is the algorithm proposed for solving the addressed problems that
online learning presents. The proposed algorithm is constructed by merging
the concepts of Primal formulation of the optimization problem, online learning
with stochastic subgradient descent solver(PEGASOS) and the support
vector reduction method SCA.