Recurring Hidden Contexts in Online Concept Learning
Stefan Mandl, Bernd Ludwig, Sebastian Schmidt, Herbert Stoyan · University of Regensburg Publication Server (University of Regensburg) · 2006
Learning systems in dynamic environments have to be able to process examples that occurred in different hidden contexts. We review several approaches to hidden context aware concept learning systems and question the appropriateness of the Calendar Apprentice domain as a real world benchmark for hidden context aware learning systems. By providing a simple lower error bound for batch learning systems in the presence of hidden contexts, we give reasons for the absence of such an agreed upon benchmark. Finally, we present a new hidden context aware concept learning algorithm, which is evaluated in a modified version of the synthetic Stagger domain.