ConGolog, Sin Trans: Compiling ConGolog into Basic Action Theories for Planning and Beyond

摘要

ConGolog is a logical programming language for agents that is defined in the situation calculus. ConGolog agent control programs were originally proposed as an alternative to planning, but have also more recently been proposed as a means of providing domain control knowledge for planning. In this paper, we present a compiler that takes a ConGolog program and produces a new basic action theory of the situation calculus whose executable situations are all and only those that are permitted by the program. The size of the resulting theory is quadratic in the size of the original program - even in the face of unbounded loops, recursion, and concurrency. The compilation is of both theoretical and practical interest. From a theoretical perspective, proving properties of ConGolog programs is simplified because reification of programs is no longer required, and the compiled theory contains fewer second-order axioms. Further, in some cases, properties can be proven by regressing the program to the initial situation, eliminating the need for a higher order theorem prove. From a practical perspective, the compilation provides the mathematical foundation for compiling ConGolog programs into classical planning problems, including, with minor restrictions, into the Plan Domain Definition Language (PDDL), which is used as the input language for most state-of-the-art planners. Moreover, Hierarchical Task Networks (HTNs), a popular planning paradigm for industrial applications can be represented as ConGolog programs and can thus now also be compiled to a classical planning problem. Such compilations are significant because they allow the best state-of-the-art planners to exploit ConGolog and HTN search control, without the need for special-purpose machinery.