Learning Opposite Concept for Machine Planning

An incomplete planning domain theory can cause an inconsistency problem in a noisy domain. To solve the problem of applying two opposite operators to the same state, we present a novel method to learn a negative precondition as control knowledge. Even though the control knowledge is unknown to a machine, it is implicitly known as opposite concept to a human. To learn the human concept, we propose a new technique to mechanically generate a graph composed of opposite operators from a domain theory and extract opposite literals. We show that the opposite concept is a special type of mutex used in Graphplan. A learned concept can simplify the operator by removing a redundant precondition while preventing inconsistencies.