ModPlan - Modern Action Planning

On the one hand - surely pushed by the series of international planning competitions - the efficiency of planning technology is continously increasing. Many recent planning systems can quickly solve rather complex planning problems. The improvement of technology is to be observed especially in suboptimal, but is also noticable in optimal planning. On the other hand, as current planning technology is still underrepresented in industrial applications, there is more transfer needed.

Henceforth, the research focus in AI planning shifts towards practical acceptance, with problem scenarios for transportation and routing, elevator scheduling, space applications, game playing, avionics, handheld setup, software verification, diagnosis in power networks, oil pipelining, etc., as indicated by the range of benchmarks currently used in planning competitions.

With recent extensions to PDDL, namely PDDL2.1, a powerful and flexible specification domain description language has been established. covering propositional and typed domain descriptions and ADL expressivity (Level 1), mixed propositional and numerical problem instances (Level 2), and temporal planning (Level 3). Additionally, PDDL2.2 provides state completion based domain axioms specified in form of derived predicates, and restricted use of exogenous events in form of timed initial predicates and action time execution windows.

For most existing planning systems, decisions that are inferred automatically can be improved by limited user guidance. Examples range from improvements to the domain encoding, the inference of domain invariances, observed goal orderings, generic types, via the choice of specialized exploration algorithms, pruning options in form of object symmetry detection, automated reduction of operator lists, learning of macro operators, to pattern database selection, hierarchical decomposition, and control rules.