A Planner Called R

SYSTEM R is a variant of the original STRIPS by Fikes and Nilsson (1971) (called forwardchaining, recursive STRIPS in Nilsson [1998]). Briefly speaking, given a conjunction of simple goals, it selects one of them to work on. If there is an action that is executable and will make this simple goal true once executed, then it does this action and then tries to achieve the other simple goals in the new situation. Otherwise, it computes a new conjunctive goal based on the actions that have this simple goal as one of their effects and recursively tries to achieve the new goal. This process continues until the original goal is achieved. The main differences between SYSTEM R and STRIPS (as given in Nilsson [1998, pp. 377–378]) are as follows: First, STRIPS maintains a global database that on starting up is the initial state description. During the planning process, it updates this global database so that it is always the description about the current situation. Thus, once STRIPS finds an action to apply, it commits to it, and no backtracking can undo this commitment. In comparison, our algorithm does not modify the initial state description. Instead, it maintains a global list Γ that represents the sequence of actions that the planner has found to this point. On backtracking, such as when reaching a dead end when pursuing a simple goal, some of the actions in Γ can be backtracked, making our planner more flexible. However, the fact that SYSTEM R maintains a global list Γ also means that checking if a fluent is true in the current situation is costlier in our algorithm than in STRIPS. Interestingly, RSTRIPS, a variant of STRIPS given in Nilsson (1980), does not change the initial situation description either and relies on regression to compute the truth value of a fluent in any future situations. Second, when attempting to achieve a simple goal, our algorithm keeps track of the sequence of all the simple goals that the planner has subgoaled to and makes sure that no cycles are involved. That is, it does not do something like “to achieve g, try to achieve g first.” This factor seemed to be crucial in making our algorithm effective on some of domains at the competition. Third, when attempting to achieve a simple goal g that cannot be made true immediately by an executable action, STRIPS selects an action that can make g true and subgoals it to the preconditions of this action. In general, some of these conditions can contain variables, and these variables are eventually grounded by matching them with the current state description. For our planner, however, the user can decide how to achieve a simple goal using domain-specific information. I describe this in more detail later in the section on controlling the planner. If there is no user-provided information, the planner uses the following strategy: It first computes the common ground preconditions of all ground actions that can make g true. If any of them is not yet true in the current situation, then the conjunction of these conditions is chosen to be the new goal. If all of them are true already, then it performs the next step. The planner then selects an action that can make g true; consider first the conjunction of all ground preconditions of this action. If this conjunction is not yet true in the current situation, then choose it as the new goal. If it is already true, then let the new goal be the conjunction of all the preconditions of this action with all the variables instantiated by some constants. Notice that in all cases, the new goal is always grounded. Articles