Simultaneous Learning and Prediction

Agents in real-world environments may have only partial access to available information, often in an arbitrary, or hard to model, manner. By reasoning with knowledge at their disposal, agents may hope to recover some missing information. By acquiring the knowledge through a process of learning, the agents may further hope to guarantee that the recovered information is indeed correct. Assuming only a black-box access to a learning process and a prediction process that are able to cope with missing information in some principled manner, we examine how the two processes should interact so that they improve their overall joint performance. We identify natural scenarios under which the interleaving of the processes is provably beneficial over their independent use.