The Loyal Opposition Comments on Plan Domain Description Languages

In this paper we take a critical look at PDDL 2.1 as designers and users of plan domain description languages. We describe planning domains that have features which are hard to model using PDDL 2.1. We then offer some suggestions on domain description language design, and describe how these suggestions make modeling our chosen domains easier. The Loyal Opposition PDDL has served as the underpinnings of the Planning Competition, and has had an enormous impact on the planning community as a whole. PDDL and STRIPS have served as a lingua franca for deterministic planning domains, making it possible for researchers to compare techniques on the same problems and enabling meaningful comparisons of those techniques. The result has been considerable progress in planning algorithms. Recently, however, the problems of interest to the planning community have changed. Goal achievement by itself is no longer enough, and the form of the goals has changed as well. Temporal planning requires meeting goals that include temporal constraints and resource constraints. While it is possible to model planning domains with temporally extended state and resources using purely propositional modeling languages, it is expensive in modeler effort and results in huge domain descriptions. By contrast, software domains such as data processing, web services integration and and information integration generally require incomplete information, incompletely known dynamic universes and sensing actions, but do not, generally, require rich models of time. In this paper, we take a critical look at the design of PDDL 2.1. Our perspective is that of researchers in planning and scheduling who have considerable experience in designing and using planning domain description languages. We will describe planning domains that pose problems for PDDL 2.1. We will then describe what we believe are the core set of features for modeling planning domains. We will show how these core features simplify the modeling of the domains of interest. PDDL and its Discontents In this section, we describe some domain models and their PDDL 2.1 representations, and then discuss reasons why the PDDL 2.1 model is problematic. Temporal Constraints and Instantaneous Events PDDL 2.1 requires that propositions hold for a non-zero amount of time before they are used as preconditions for actions. However, PDDL’s underlying representation of states is not based on intervals, and so this makes it difficult to specify the interaction between multiple events that modify the same proposition. It will also be difficult to see what domain axioms resulted in the separation of two events that modify the same proposition in the final plan. Part of the reason for this is that sometimes actions require additional constraints to decide whether the action sequence is legal, and PDDL 2.1’s semantics forbid such plans when there is a possibility of plan execution failure. The following example appears in (Fox & Long 2003) . The model includes action A with preconditions P ∨Q and effect R, and action B with preconditions P and effect ¬P and S. These two actions are considered mutually exclusive because a plan in which A and B execute concurrently may fail. We might have A and B execute at the same time in a state where ¬Q holds, for example. PDDL assumes that the executions of A and B are actually ordered, and that the order is arbitrary. Thus, if the state happens to be one where ¬Q holds, it may be thatB happens first, makingA fail both its preconditions. Modeling this scenario by asserting the intervals of time over which P and Q hold, and forcing actions to declare how long the precondition must hold, clarifies the situation. For example, if A requires P to hold for 1 unit of time prior to A’s execution, then in order for the plan to be legal, B can’t happen at a time that makes P true for less than a unit of time. This makes it possible to both post and check the constraints that must hold for both A and B to happen concurrently. If P has held for long enough, then A and B can happen at the same instant, even in a state where ¬Q holds. A related problem is that it is hard to express certain temporal constraints in PDDL 2.1. For example, suppose that an action Take-Picture requires a proposition stable to hold for a period lasting from at least 5 seconds before to at least 5 seconds after the Take-Picture ends. Furthermore, the camera expends energy only during the time of the exposure, which lasts 24 seconds. During the intervening 5 seconds before and after the exposure, the camera could be involved in other activities such as changing the filters, but can’t slew (as this results in vibration). PDDL 2.1 does not allow direct expression of the constraints on stability. As we said previously, there is no way