Learning to reason in a Probably Approximately Correct manner

This paper investigates the development of a knowledge base (KB) of logical functions, that can be used to do reasoning, from the consolidation of training examples of those logical functions. The work is based on the L2R (Learning to Reason) framework. A L2R agent only needs to answer knowledge queries that are relevant to its environment in a Probably Approximately Correct sense. We develop an L2R KB of Boolean functions by training a context-sensitive Multiple Task Learning network on examples of the truth tables of those functions. Reasoning is abstracted as a deduction task of determining whether a query Q is entailed by the KB. This is done by testing the neural network model on the truth table examples of Q to determine if the L2R KB agrees. Experimental results show that for different logical KBs and deduction rules the L2R approach shows promise.