Partial Compilation of Strategic Knowledge1

Many system building efforts in artificial intelligence intentionally begin with expressively rich and flexible declarative structures for the control of problem solving-especially when the best problem solving strategies are not known. However, as experience with a system increases, it sometimes becomes desirable to compile declarative knowledge into procedures for purposes of efficiency. We present a paradigm for compilation which begins with declarative opportunism, moves to a phase of heuristic implementation of a partial plan and finally evolves into a fully elaborated procedure. We use the PROTEAN geometric constraint satisfaction system as an example. Using results from a purely declarative structure, we were able to compile strategic knowledge into a procedure for planning a solution. The problem ior of the new system is reported. solving behavKnowledge compilation offers an engineering solution to the problem of combining the flexibility of a declarative representation of knowledge with the efficiency of a more procedural representation. For applications in which knowledge is changing frequently, the benefits of declarative representations may outweigh considerations of efficiency (especially during development). For others, in which run-time efficiency is more important, the use of declarative representations for initial development must be followed by compilation of knowledge. As knowledge-based systems become larger, there is increasing use of separate meta-level knowledge structures (sometimes called strategic or control knowledge) to reduce the complexity and increase the understandability of these systems [Davis, 1980, Clancey, 1985, Hayes-Roth, 1985, Hewitt, 1972, McDermott, 19781. In this paper we show the results of compiling parts of this strategic knowledge, represented declaratively, into a partial plan that instantiates major control decisions. It is useful to have a rational method by which tion from declarative to procedural forms of strategic can be made gracefully. In this paper, we argue: the transiknowledge ‘This work was funded in part by the following contracts and grants: NIH GM07365, DARPA N00039-83-C-0136, DARPA N00039-86-C-0033, NIH RR-00785, NASA-Ames NCC-2-274, Boeing Computer Services W271799, and a gift from Lockheed Corp. We would like to thank Alan Garvey, Craig Cornelius and Barbara HayesRoth for discussion of their experimental results. We also thank Jim Brinkley, Bruce Duncan, John Brugge and Oleg Jardetzky for collaborative research on PROTEAN. 1. That the separation between strategic knowledge and domain knowledge (as in the PROTEAN/BBl blackboard system) is useful in the development of efficient problem solving strategies. We suggest a three stage paradigm with which this development can usefully be viewed. 2. That if strategic knowledge is represented declaratively and separated from domain problem solving knowledge, then compilation of strategic knowledge can be performed and integrated within domain problem solving knowledge. 3. That the compilation of parts (but not all) of the problem -solving knowledge yields plans in which flexibility is sacrificed for efficiency. These plans embody a set of decisions that may anticipate global problem solving strategy better than more locally focussed strategy knowledge. We substantiate these claims with examples from the PROTEAN system for the determination of protein structure [Altman and Jardetzky, 1986, Brinkley et al., 1986, Hayes-Roth et al., 1986131. PROTEAN is a geometric constraint satisfaction system described in section II. In one version of this system, Hayes-Roth and coworkers have shown that declarative control structures can be used to control reasoning about constraint satisfaction in spatial assembly problems [Hayes-Roth et al., 1986a]. We have compiled elements of this strategic knowledge and have been able to implement plans to guide problem solving without any modification of the basic domain problem solving actions. The plans prescribe partial sequences of actions. Portions of the problem solving for which there is no plan prescription are controlled opportunistically with heuristics. A. Three Phases of evelspment for Procedures The development of a computational procedure for the solution of difficult problems can be usefully divided into three phases. The first phase, which we calf the opportunistic phase, is characterized by the use of architectural frameworks in which there is considerable freedom for a designer to experiment with different formulations of the search and different strategies for controlling it. Having gained experience from work within the opportunistic phase of development, we can enter the partial plan phase. A partial plan provides an incomplete specification of the actions required for a solution. In this phase a designer draws upon the heuristics and experience gained during experimentation to increase efficiency with a more rigid problem solving control plan. These plans are not fully prescriptive for problem solving, however, and some of the declarative strategic knowledge may remain when the plan prescribes nothing. Altman and Buchanan 399 From: AAAI-87 Proceedings. Copyright ©1987, AAAI (www.aaai.org). All rights reserved.