The current emphasis of the research in learning theory is on the study of inductive learning (from exarples) of concepts (binary classifications of examples). The work in AI identifies other tasks, such as reasoning, as essential for intelligent agents, but those are not supported by the current learning models. The Learning to Reason framework was devised to reconcile inductive learning and efficient reasoning. The framework highlights the fact that new learning questions arise when learning in order to reason. This paper addresses the task of deductive reasoning, and investigates learning to reason problems in which the examples seen are only partially specified.

The paper presents several interpretations for partial information in the interface with the environment, and develops model based representations and reasoning algorithms that are suitable to deal with partially observable worlds. Then, learning to reason algorithms that cope with partial information are developed. These results exhibit a tradeoff between learnability, the strength of the oracles used in the interface and the expressiveness of the queries asked.

This work shows that one can learn to reason with respect to expressive worlds, that cannot be learned efficiently in the traditional learning framework and do not support efficient reasoning in the traditional reasoning framework.