In the beginning
Paul Compton at UNSW proposed Ripple-Down Rules (RDR) in the late 1980s. Paul published his early RDR papers, Knowledge in context: a strategy for expert system maintenance, at the Second Australian joint conference on Artificial intelligence (AI 88) in 1988 and A philosophical basis for knowledge acquisition, at the Third European Workshop on Knowledge Acquisition for Knowledge-Based Systems (EKAW 89) in 1989. Most foundations of RDR are discussed in these two papers. If you properly understand RDR approach and get to know unique features of RDR, you should read these papers.
What we learn from these papers? At least I think we can
- understand problem on which RDR focuses;
- learn about philosophical basis; and
- understand basic structure of RDR-based systems.
First thing that we notice is that RDR approach expects human for acquiring knowledge. This differentiate RDR from machine learning (ML), which usually does use data to learn its model. However, it is also necessary to remember that we don't need to choose either RDR or ML approach, since they can be complement each other. Usually data used by ML approaches need label that specifies each case's class. The class might be labelled without intervening human, but most cases it is essential for the human to involve in the labeling process. I'll discuss this issue later, but this time you need to know this fact - RDR and ML are complement each other. We can also find that RDR approach focuses on how to maintain human knowledge in a system. Like human, who continually learns over his/her whole life, RDR assumes that the systems that behave based on their encoded knowledge should continually acquire knowledge. This is significant in the practical expert systems. A well known example is R1 expert system developed by John McDermott. Its initial system was established in 1981 with 777 rules, but four years later the number of rules increased 3303 (McDermott 1981, Bachant and McDermott 1984). Early RDR researchers experienced similar problem when developing GARVAN-ES1. The following figure how rule base size increase after the system was routinely used (Compton and Jansen 1990). RDR approach was developed keeping in mind the fact that rules everlastingly increase over the time.
Why rules are ever increasing? To explain this issue, early RDR researchers addressed knowledge in "context"(Compton and Jansen 1988). Compton and Jansen nicely summarized RDR's view on knowledge and its everlasting increase over the time as follows:
Knowledge engineering, obtaining knowledge from experts and incorporating it into expert systems is difficult and time consuming. We suggest that these difficulties arise because experts never report on how they reach a decision, rather they justify why the decision is correct. These justifications vary markedly with the context in which they are required, but in context they are accurate and adequate; the difficulties in knowledge engineering arise from taking the justification out of context. We therefore hypothesized that knowledge engineering may be obviated, particularly in the long term maintenance of expert systems, if the rules experts provide are used in the context in which they are given.
Learn about philosophical basis.
In extension of the concept of knowledge in "context", Compton and Jansen wrote A philosophical basis for knowledge acquisition (Compton and Jansen 1989), the most sited paper in RDR community. Even though they admitted that "Knowledge acquisition is a purely practical problem to be solved by experiment, independent of philosophy", they thought that they should write the paper that accounts for philosophical difference of their new approach, since "the experiments one chooses to conduct will be influenced by one's implicit or explicit philosophy of knowledge, particularly if this philosophy is taken as axiomatic rather than as an hypothesis."
In conjunction with the fact of knowledge in context, they criticized the physical symbol hypothesis that implies some underlying knowledge structure can be found is a misleading philosophical basis for knowledge acquisition and representation. They thought "the 'insight' hypothesis of B. Lonergan better explain the flexibility and relativity of knowledge and may provide a more suitable philosophical environment for developing knowledge acquisition and representation tools." More details on philosophical basis, please read (Compton and Jansen 1989).
Understand basic structure of RDR-based systems.
The knowledge engineer is the person responsible for placing the knowledge into the expert system’s knowledge base using rule adjuster. In conventional expert systems the knowledge engineer is the interface between the human expert and the expert systems.
The user is anyone who will be using the expert systems as a decision-making aid. The expert systems is ultimately intended for the benefits of the user, not for that of the knowledge engineer or the domain expert. The interface handles all inputs to the computer and controls and formats all output.
The inference engine examines the status of the knowledge base and working memory in order to determine what facts are known at any given time, and to add any new facts that become available. In addition, it provides for the control of the session by determining the order in which inferences are made.
The knowledge base contains two types of knowledge, that is, facts and rules. The facts within a knowledge base represent various aspects of a specific domain that are known prior to the exercise of the expert system. The rules within the knowledge base are heuristics that are developed through intuition, experience, and judgement. Heuristics are intended for use in obtaining acceptable solutions, no best or optimal results. The rules represent domain expert’s heuristics constructed based on its perception of the knowledge engineer.
The working memory consists of facts that are determined for the specific problem under considering during (and at the conclusion of) the consultation session. More specifically, the results of the inference process are new facts and these facts are stored in the working memory. The rule adjuster serves as a rule editor. It may allow for various checks on rules and support incorporation of learning into expert systems.
References
McDermott, J.: R1: the formative years, in Readings from the AI magazine. 1981, American Association for Artificial Intelligence. pp. 93-101.
Bachant, J. and J. McDermott: R1 Revisited: four years in the trenches, in Readings from the AI magazine. 1984, American Association for Artificial Intelligence. pp. 177-188.
Compton, P. and R. Jansen: A philosophical basis for knowledge acquisition. Knowledge Acquisition. 2(3), pp. 241-258 (1990).
Compton, P. and R. Jansen: Knowledge in context: a strategy for expert system maintenance, in Proceedings of the second Australian joint conference on Artificial intelligence. 1988, Springer-Verlag New York, Inc.: Adelaide, Australia. pp. 292-306.
No comments:
Post a Comment