Computational Trade-offs in Experience-Based Reasoning

Navigational p th planning isa classical problem in robotics. Traditional approaches u e goal-directed heuristic search of problem spaces defined by spatial models of the navigation world. Case-based reasoning offers an alternative approach. In the Router project, we have combined the case-based method with the model-based method. Since Router is a multistrategy system, it provides anexperimental estbed to study some of the hypotheses ofcase-based reasoning. In this paper, we report on a set of experiments that examine four hypotheses: (i) the case-based method more efficient than the model-based m thod, (il) the case-based method produces plans of quality equal to those produced bythe model-based method, (iii) the case-based method requires less knowledge but has the same problem-solving coverage asthe model-based method, and (iv) cases need to be decomposed into partial cases for efficient and effective problem solving. We find that while hypothesis (i) is true, the others are questionable. Goals and Motivations NavigationM path planning is a classicM problem in autonomous mobile robotics. Qualitatively, a good method for robot planning would obey the following constraints: (1) since robots have access only to bounded computational resources, the planning method would require only limited processing and memory, (2) since robots need to perform in close real time, it would Ca) successfully form accurate plans and (b) form them very efficiently; and (3) since robots often operate in dynamic worlds, it would not assume complete and correct knowledge of the world. The contradictions between constraints 1 and 2, 2 and 3, 2(a) and 2(b), make navigational path planning a challenging problem for AI. Most AI approaches to navigational planning use goal-directed heuristic search of problem spaces defined by spatial models of the navigation world [Fikes, Hart, and Nilsson 1972; Kuipers and Levitt 1988; McDermott and Davis 1984]. Since it employs a spatial model, we will refer to this family of methods as modelbased planning. Ifthe robot’s navigation world is static and the robot planner has complete and correct knowledge of this world, then model-based planning guarantees both efficient processing and high-quality solutions [Aho et. a11974]. However, complete world models are impossible to provide for operation i a dynamic environment, and current limitations of robot sensors and learning methods generally make it difficult toacquire such knowledge directly. Experience-based reasoning [Hammond 1989, Kolodner and Simpson 1989, Riesbeck and Schank 1989, Sussman 1975, Winston 1982] offers an alternative approach to model-based navigational path planning. In this approach, the robot planner euses previously formed plans to solve new planning problems: given a planning problem, the planner etrieves a past case of planning from its memory and adapts the plan stored in the case to meet the specifications f the current problem. This case-based family of methods promises several advantages over the model-based family [Kolodner 1993]: (i) since it relies on reusing specific experiences for solving new problems rather than reasoning from a general model of the navigation world, it provides for more fficient planning, and (ii) since can start with relatively few cases in memory and dynamicMly acquire new cases based on the reasoner’s interactions with the world, it makes few assumptions about the completeness and correctness of world knowledge. Note that, in theory, case-based planning offers these benefits without incurring any significant loss in the quality of solutions produced. In the Router project, we have combined the casebased method with the model-based method for navigationM planning. Since Router is a multistrategy system, it provides an experimental testbed for studying some trade-offs in multistrategy reasoning, and also for studying some of the basic hyptheses of case-based reasoning. In this paper, we report on a set of experiments that examine four hypotheses: (i) the case-based method is more efficient than the model-based method, (ii) the case-based method produces plans of quality equal to those produced by the model-based method, (iii) the case-based method requires less knowledge but 55 From: AAAI Technical Report WS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. has the same problem-solving coverage as the modelbased method, and (iv) cases need to be decomposed into partial cases for efficient and effective problem solving. We find that while hypothesis (i) is true, the others are questionable. The goal of this paper is to report on these experiments, draw some lessons on the utility of the case-based and model-based methods for navigational planning, and the appropriateness of a specific framework for multistrategy reasoning. General System Design Router is a goal-directed multistrategy navigational path planning system. It operates in two kinds of navigation worlds: representations of office buildings such as the College of Computing Building (CoC) Georgia Tech (GT), and representations of urban areas such as the Georgia Tech campus in Atlanta. In both worlds, the input to Router is a pair of spatial locations representing the initial and goal positions that the path should connect. The initial and goal locations are among the intersections between the pathways in the world; the pathways are the streets in the Georgia Tech domain and the corridors and hallways in the College of Computing domain. The output produced by the system is an ordered set of path segments (streets, hallways) between the initial and the goal locations. In addition, Router accepts as input feedback on the execution on a plan it produces and attempts to learn from both its successes and failures. Router combines the model-based and case-based methods for planning navigation paths in these domains. It integrates these two methods in three dimensions: planning, memory, and learning. In the following subsections we briefly characterize the main components of Router in these three dimensions.