Workshop Summaries Report on the IJCAI-95 Workshop on Anytime Algorithms and Deliberation Scheduling Montreal, Canada - August 20, 1995 Michael Pittarelli Computer Science Department SUNY Institute of Technology Utica, NY mike @ cs.sunyit.edu This workshop was organized to bring together workers in some closely related areas of artificial intelligence research that have seen a sharp increase in activity in recent years. The three invited speakers, Eric Horvitz (Microsoft Research), Thomas Dean (Brown Univ.), and Stuart Russell (UC-Berkeley), are each pioneers in the field, which can be loosely characterized as the study and construction of agents that exhibit metalevel rationality. One goal of the workshop was to clarify some of the terminology introduced by various research groups starting in the late '80s. The term "anytime algorithm" was coined by Dean and Boddy [1] to apply to methods that provide approximate answers to difficult problems in such a way that, minimally: (1) An answer is available at any point in the execution of the algorithm (2) The quality of the answer improves with an increase in execution time. Additional properties may include the following: (3) Preemptibility (4) Continuity of the function from time to quality (5) Diminishing marginal improvement of quality with time. "Deliberation scheduling" is the term applied by Dean and Boddy to the process of determining the optimal amount of time to allocate to deliberation (e.g., planning, scheduling) before acting (with a plan or schedule that might be regarded as suboptimal, if deliberation costs are not taken into account). Horvitz [2] deals with computational resource costs that may involve more than just computing time and introduces the term "flexible algorithm." He also introduces the notion of "bounded optimality", which takes into account the resources available to an agent. This concept is extended to asymptotic bounded optimality by Russell and Subramanian [ref95]. The invited speakers each recognized the contributions of the others and also provided some historical background, e.g., the work of I.J. Good on metalevel rationality ("Type II rationality"), Herbert Simon on bounded rationality, and Ronald Howard on the value of information. There was also mention of related work in other areas of computer science: real-time systems, numerical computing, operating systems, and database systems. There were 13 contributed papers, each of which is available via anonymous ftp to cs.sunyit.edu (in the "anytime" directory). Unfortunately, the invited talks do not correspond to any of the submitted papers. A special issue of the SIGART Bulletin containing papers by the three invited speakers plus selected workshop submissions is planned for early 1996. In the area of deliberation scheduling, papers were contributed by Scott Anderson and Paul Cohen ("The duration of deliberation") and by Yagil Ronen and Martha Pollack ("Value-density algorithms for deliberation scheduling"). Anderson and Cohen explore solutions to the problems involved in measuring actual deliberation time: dependence on a particular implementation; interference caused by adding code for monitoring purposes; replicability of results; etc. Ronen and Pollack discuss the applicability to deliberation scheduling of algorithms that have been developed for process scheduling in real-time operating systems. Performance monitoring and composition are discussed in three of the papers: "Composition and monitoring of anytime algorithms" (Shlomo Zilberstein), "Monitoring the progress of anytime algorithms" (EricHansen and Shlomo Zilberstein), and "Programming with anytime algorithms" (Joshua Grass and Zilberstein). Four papers discuss applications to planning and scheduling problems: "An adaptive real-time agent for resource planning within the context of naval anti-air warfare" (Jean Berger and F. Loranger), "Design to time scheduling and anytime algorithms" (Alan Garvey and Victor Lesser), "k-best: a new algorithm for making real-time decisions" (Joseph Pemberton), and "Focusing attention in anytime decision-theoretic planning" (Peter Haddawy). Chao-Lin Liu and Michael Wellman ("On state-space abstraction for anytime evaluation of Bayesian networks") discuss approximate evaluation of Bayesian networks based on incrementally refining statespaces. Marco Ramoni ("Anytime influence diagrams") discusses anytime decision making with interval-valued influence diagrams. Also contributed were "Anytime algorithms for constraint satisfaction and SAT problems" (Richard Wallace and Eugene Freuder) and "Reactive/anytime agents: towards intelligent agents with real-time performance" (Martin Adelantado and S. De Givry). Due in part to the success of the workshop, a proposal for a AAAI Fall Symposium (1996) on "Flexible computation in intelligent systems" was recently approved. References [1] T. L. Dean and M. Boddy. An analysis of time-dependent planning. In Proc. of AAAI-88, St. Paul, 1988, pp. 49-54. [2] E. Horvitz. Reasoning about beliefs and actions under computational resource constraints. Knowledge Systems Laboratory Report KSL-87-29, Stanford University, 1987. Also in L. Kanal et al., Eds. Uncertainty in Artificial Intelligence 3. North Holland, 1987 S I G A R T Bulletin, Vol. 6 No. 4

/lp/association-for-computing-machinery/report-on-the-ijcai-95-workshop-on-anytime-algorithms-and-deliberation-wf1J5ZNKV5