Representing and reasoning about temporal information are of interest in artificial intelligence. The most prominent research in ths area has been formal. The wealth of competing formalisms [Shoham, 1987; Allen, 1984; Bacchus et al., 1981] provides a number of choices for designing a model of time and for representing and reasoning about temporal intervals; temporal positions can be reified or nonreified; and temporal reasoning can involve the scope of first-order predicate logic or can be restricted to the Horn clauses of a PROLOG-like system. When we implement a knowledge-based system that uses temporal representation and temporal reasoning, the functionalities of the application and the nature of domain information must suggest the appropriate temporal model. In this paper, we present a class of applications and an implementation strategy for temporal reasoning that is appropriate for that class. Our application involves a reactive planning system in the domain of medical therapy. As in many complex real-life domains, the physiological processes that medical therapy seeks to control are too complex to be modeled completely. Instead, medical experts start with a standard plan for solving a class of problems, and then use domain-specific heuristics to modify the standard plan in response to information gathered during plan execution. This type of planning relies on the planner specializing and instantiating a skeletal plan that specifies procedures of actions that are known to be effective in similar cases. The system assumes the interleaving of generation and execution of plan steps. The system initially follows the procedues specified in the standard skeletal plan, and, as the plan is executed, modifies the standard plan periodically in reponse to the observed patient reactions to the earlier actions. We call this problem-solving method episodic skeletal-plan refinement (ESPR) [Tu et al., 1987].
