Planning with Independent Task Networks

Abstract

Task networks are a powerful tool for AI planning. Classical approaches like forward STN planning and SHOP typically devise non-deterministic algorithms that can be operationalized using classical graph search techniques such as A*. For two reasons, however, this strategy is sometimes inefficient. First, identical tasks might be resolved several times within the search process, i.e., the same subproblem is solved repeatedly instead of being reused. Second, large parts of the search space might be redundant if some of the objects in the planning domain are substitutable. In this paper, we present an extension of simple task networks that avoid these problems and enable a much more efficient planning process. Our main innovation is the creation of new constants during planning combined with AND-OR-graph search. To demonstrate the advantages of these techniques, we present a case study in the field of automated service composition, in which search space reductions of several magnitudes can be achieved.