Heuristics for Planning in Domains with Unbounded, Incomplete Information

While the Internet Softbots project (Etzioni & Weld 1994) investigated the use of planning algorithms to control software agents (Weld 1996), until recently the area has seen little activity. However, the excitement surrounding the Semantic Web, has led to renewed interest in software agents, web service composition, and related topics. With the development of DAML-S (Burstein et al. 2002; Ankolenkar et al. 2001), a DAML+OIL ontology for describing properties and capabilities of Web services in the form of inputs (similar to preconditions) and outputs (effects), an infrastructure has been created where agents can automatically discoverWeb Services that are modeled as actions, and then create and execute plans (McIlraith, Son, & Zeng 2001; McDermott 2002). The major challenge confronting a planning agent operating in an environment such as the Internet is uncertainty about the world-state. But most prior work on planning under uncertainty (e.g., (Peot & Smith 1992; Kushmerick, Hanks, & Weld 1995; Pryor & Collins 1996; Weld, Anderson, & Smith 1998; Bonet & Geffner 2000; Bertoli, Cimatti, & Roveri 2001)) is inadequate because of scalability issues. When a single world-state is as complex as a UNIX file system or the Internet, representing belief states explicitly is out of the question. Furthermore, one requires a sophisticated language simply to represent sensing actions, which like UNIX ls, can return an unbounded amount of information at execution time. Clearly the closed world assumption (Reiter 1978) is no longer valid, yet the open-world assumption leads to paralysis. In this case the agent must execute sensing actions to discover objects and relations while it plans. Furthermore, the agent must maintain a record of where it has local closed world (LCW) information (Etzioni, Golden, & Weld 1994; 1997).