The Model Checking Integrated Planning System (MIPS)

With the Model Checking Integrated Planning System MIPS, model checking has eventually approached classical AI planning. It was the first planning system based on formal verification techniques that turned out to be competitive with the various Graphplanor SAT-based approaches on a broad spectrum of domains. MIPS uses binary decision diagrams (BDDs, introduced by Bryant (1986)) to compactly store and operate on sets of states. More precisely, it applies reduced ordeded binary decision diagrams, which we will refer to simply as BDDs for the rest of this article. Its main strength compared to other, similar approaches lies in its precompiling phase, which infers a concise state representation by exhibiting knowledge that is implicit in the description of the planning domain (Edelkamp & Helmert 1999). This representation is then used to carry out an accurate reachability analysis without necessarily encountering exponential explosion of the representation. The original version of MIPS, presented at ECP99, was capable of handling the STRIPS subset of PDDL. It was later extended to handle some important features of ADL, namely domain constants, types, negative preconditions and universally quantified conditional effects. Other extensions include two additional search engines based on heuristics, one incorporating a single-state hillclimbing technique very similar to Hoffmann’s FF, the other one making use of BDD techniques, thus combining heuristic search with symbolic representations. However, as the former does not contribute many new ideas, its merits mainly lying in the combination of Hoffmann’s heuristic estimate with the preprocessing techniques of MIPS, we won’t dwell on it. Neither will we say much about the symbolic heuristic search techniques included in MIPS, namely the BDDA* and Pure BDDA* algorithms, as those were disabled in the AIPS 2000 planning competition in favor of the original MIPS planning algorithm, partly because it turned out to perform better on some domains, partly because it always yields optimal (sequential) plans, which we consider an important property of the planner that counterbalances some of its weaknesses in performance compared to other current planning systems such as FF. Readers interested in those