Planning with Incomplete Information in the UNIX Domain
Oren Etzioni, Neal B. Lesh · 1993
This paper serves two goals. First, we advocate real-world software environments, such as operating sys-tems or databases, as domains for planning research. Second, we describe our ongoing research on planning with incomplete information. In both cases, we fo-cus on the UNIX planning domain as a testbed for our implementation and a source of concrete examples and intuitions. l%eal-world software environments, such as UNIX, have a number of attractive features as testbeds for plan-ning research. First, it is relatively easy to develop "embodied agents, " which interact directly with a soft-ware environment by issuing commands and interpret-ing the environment’s feedback [6]. Such agents facil-itate testing planners experimentally and integrating planning with execution and error recovery capabili-ties. Second, software environments circumvent many thorny, but peripheral, research issues that have to be addressed in physical environments (e.g., overcoming sensor noise, representing liquids, shapes, etc.). Third, software experiments are relatively easy to perform, control, and repeat, facilitating systematic experimen-tal research of the sort advocated by [10] and oth-ers. Finally, software facilitates the dissemination and replication of research results. It is straightforward to distribute multiple copies of a software agent, and many software environments (e.g., UNIX) are easily accessible to most researchers. Simulated worlds share some of these features but, software environments are readily available (sophisti-cated simulations can take years to develop and per-fect). Furthermore, software environments are real. *We thank Richard Segal, our co-softboticist, for his numerous contributions to this research. Thanks are also