Reliable localized on-line learning in non-stationary environments

On-line learning allows to adapt to changing nonstationary environments. But typically with on-line learning a hypothesis of the data relation is adapted based on a stream of single local training examples, continuously changing the global input-output relation. Hence with these single examples the whole hypothesis is revised incrementally, which might be harmful to the overall predictive quality of the learned model. Nevertheless, for a reliable adaptation, the learned model must yield good predictions in every step. Therefor, the IRMA approach to online learning enables an adaptation that reliably incorporates a new example with a stringent local, but minimal global influence on the input-output relation. The main contribution of this paper is twofold. First, it presents an extension of IRMA regarding the setup of the stiffness, i.e. its hyper-parameter. Second, the IRMA approach is investigated for the first time on a non-trivial realworld application in a non-stationary environment. It is compared with state of the art algorithms on predicting future electric loads in a power grid where a continuous adaptation is necessary to adapt to season and weather conditions. The results show that the performance is increased significantly by IRMA.