The Research On Extending Trajectory GraphPlan With Probabilistic PDDL Subset

Intelligent planning is one of the most important research fields in AI. Classicalplanning problem supposes the information about the world is totally complete and whenan action is performed, the effects of the action are definite, which limits the classicalplanning algorithms only to solve small-scaled model planning problems. To solve realproblems, nowadays, people have attached importance to the research on incompleteinformation and uncertain effects. Probabilistic planning has extremely aroused attention because it can quantitativelydescribe uncertain information accurately. Blum & Furst developed TGP (TrajectoryGraphPlan), which extended GraphPlan that can only solve classical planning problemswith STRIPS to solve probabilistic problem domain with the effects of an operatorhaving different probabilistic outcomes. Relevant results have shown TGP is faster thanany other planner that can also solve the same planning problems. Conditional effects of actions can be also seen as an approach to describe uncertaineffects of an operator. But it is different from effects with probabilistic outcomes.Probabilistic outcomes can be defined by means of experience in advance. When anaction is performed, it can cause a probabilistic outcome and only one outcome. So youcan see various probabilistic outcomes of an action as mutually exclusive outcomes.While whether conditional effects happen depends on which state the action is in.conditional effects of an action share the preconditions of the action, but have their owneffect conditions. These conditional effects may co-exist in one state if their conditionsare true and are non-exclusive. It is very valuable to integrate these two approaches, andthat is why we propose this paper. In this paper, we first review the development of intelligent planning and planninglanguage. Second, we describe existing approaches dealing with conditional effects. Then,we analysis probabilistic planning algorithms and at last we describe how to extend TGPwith PPDDL (Probabilistic Planning Domain Definition Language) subsets to solve moreplanning problems.