Planning for Temporally Extended Goals

this paper appears in Proceedings of AAAI '96, pp. 1215-1222. F. Bacchus and F. Kabanza / Temporally Extended Goals 2 is, of course, necessary if the agent is to be able to accomplish a range of tasks under varying conditions. Yet this flexibility also poses a problem: how do we communicate to such an agent the task we want accomplished in a sufficiently precise manner so that it does what we really want. In the area of planning, methods and algorithms are studied by which, given information about the current situation, an intelligent agent can compose its primitive abilities so as to accomplish a desired task or goal. The afore mentioned problem then becomes the problem of designing sufficiently expressive and precise ways of specifying goals. Much of the work in planning has dealt with goals specified as conditions on a final state. For example, we might specify a goal as a list of literals. The intent of such goals is that the agent should find a plan that will transform the current situation to a configuration that satisfies all of the literals in the goal. Any plan that achieves such a satisfying final state is deemed to be correct. However, there are two major limitations in the expressive power of such goals. First, not all goals are "final state" goals. For example, we might have goals of maintenance or reaction, where the agent must try to maintain a condition or respond within a limited time frame to a condition. Such goals can be important, but they cannot be specified as commands to reach a certain final state. Second, there are many important constraints we might wish to place on the agent's behavior that similarly cannot be expressed using final state semantics for goals. For example, if the agent's goal is to minimize disk usage, we might want to constr..