An Approach to Dependency Directed Backtracking using Domain Specific Knowledge

Abstract

We describe a problem solver that met non-chronological backtracking in situations involving tradeoffs. The novel aspect of the problem solver is it ability to weigh advantages and disadvantages of alternatives at choke points. Whenever untenable situations arise, this information is available to the backtracker to determine the most appropriate backtracking point. By endowing the backtracker with access to domain-specific knowledge, a highly contextual approach to reasoning in dependency directed backtracking situations can be achieved. An area of investigation now commonly referred to as "Dependency Directed Reasoning," has had a major impact on AI research in the last decade (de Kleer et.al, 1977; Doyle, 1978, 1979, 1980; McAllcster, 1980; Stallman and Sussman, 1977). In this paper we describe some dependency directed reasoning features of a problem solver called PLANET (Dhar, 1984). It has been designed to help planning managers in a large computer manufacturing company with the formulation and investigation of models for allocation of resources. Since the process of resource planning 1 involves making assumptions that are continually subject to revision, dependency information plays a crucial role in the maintenance and incremental change of planning models. What is of particular interest in this paper is a heuristic procedure for dependency directed backtracking that addresses one drawback of existing dependency frameworks, namely, determining what belief (set of assumptions) in an existing model to change whenever an undesirable state arises. 2. The Need far Dependency Directed Backtracking The problem of formulating a resource planning model has two important features. First, as assumptions about various parts of the task environment are made, choices in other parts of the environment are constrained. Second, there are usually resource requirement tradeoffs among the alternatives that can be made. These features 1 We use the term paw and planning to refer to busbmt (manufacturing or resource) plan(ing), and not a plan as normally understood in AI. A manufacturing plan it an interrelated set of choices about various parts of the manufacturing process. A resource plan indicates the types and amounts of resources a manufacturing plan requires.