Many planning systems meant for single-agent problem solving have had to deal with unanticipated events. For example, PRS [Georgeff &Lansky, 1987][Georgeff &Ingrand, 1988] and MEDIC [Turner, 1994][Turner, 1989] have both explored the problem of events occurring during the execution of a single agent's plan. PRS, for example, uses malfunction procedures and meta-knowledge to handle failures that have been anticipated and domain specific meta-knowledge to reason about multiple unrelated failures; the development of a general solution for reasoning about unexpected (unrelated) failures is left for future work. MEDIC uses contextual schemas to provide information about an unanticipated event's importance in a particular situation and about how to handle the event appropriately. ORCA [Turner &Stevenson, 1991][Turner, 1994], which is based partly on MEDIC, is extending MEDIC's event-handling mechanism to cope with real-world domains; the work reported here will ultimately be folded into ORCA.
Others have looked at the problem of handling events in multiagent situations. For example, Phoenix [Cohen et al., 1989], a multiagent system operating in the firefighting domain, has agents that interleave planning and execution, are capable of reactive responses to anticipated events, and can perform deliberative planning. However, unanticipated events are not handled by individual agents, which have only a local view of the problem-solving situation, but rather by a central fire-boss with a global view of the situation. Individual agents do not reason about the effect their actions may have on others, and the system uses a set of predefined events a single agent may encounter. Work on Partial Global Planning (PGP) [e.g.,]pgp also deals with events in a distributed AI environment. However, the agents we are interested in are more autonomous than the typical PGP agents; consequently, they have more flexibility to handle events on their own. Cammarata et al. (1983) were concerned with agents that are semi-autonomous (airplanes) in their air traffic control system. However, they dealt with a relatively impoverished set of event types, and relied on planning by a single agent in most cases to handle the event.