Loading...
Please wait, while we are loading the content...
Similar Documents
ECLIPS: An extended CLIPS for backward chaining and goal-directed reasoning
| Content Provider | Semantic Scholar |
|---|---|
| Author | Homeier, Peter V. Le, Thach Cam |
| Copyright Year | 1991 |
| Abstract | Realistic production systems require an integrated combination of forward and backward reasoning to reflect appropriately the processes of natural human expert reasoning. A control mechanism that consists solely of forward reasoning is not an effective way to promptly focus the system’s attention as calculation proceeds. Very often expert system programmers will attempt to compensate for this lack by using data to enforce the desired goal-directed control structure. This approach is inherently flawed in that it is attempting to use data to fulfil the role of control. This paper will describe our implementation of backward chaining in CLIPS, and show how this has shortened and simplified various CLIPS programs. This work was done at the Aerospace Corporation, and has general applicability. 1. DESIGN CONSIDERATIONS The Aerospace Corporation has been using expert system technology since the mid-1980s, beginning with a system to diagnose anomalies in the attitude control system of the DSCS III satellite. These experiments showed the value of expert system technology in Air Force programs, and identified key special requirements. The Portable Inference Engine (PIE) project (Le and Homeier 1988) was intended to produce a single language and environment for Air Force expert systems to be written and run across a wide variety of hardware bases. CLIPS was identified as meeting most of these requirements. Singular among these is real-time response (Laffey et al. 1988), which we interpret in the context of expert systems as time efficiency. The Rete net algorithm is known to possess optimal efficiency for matching many patterns to many objects (Forgy 1982). It is based on the assumption of a slowly changing state. Air Force requirements involve high rates of data to be processed in real time. Thus the state is changing rapidly, perhaps completely in a short time. This condition does not satisfy the stated assumptions of the Rete net, and thus adaptations of the algorithm are needed. Typically, 90% of the execution of an expert system is spent in matching (Gupta 1985). Our approach to improving the speed of matching is to reduce the number of rules being considered for matching at any one time. Most real expert systems do not have a flat structure, where all rules are expected to be ready to fire at any point (Winston 1984). Rather, in many cases, there is effectively a current focus of attention, where a few rules are doing the work for the moment, and the rest of the expert system is essentially waiting around for its turn to contribute to the task. Although that waiting sounds passive, it is truly active, since the left-hand-sides of those rules are participating in the Rete net matching process, and all facts that apply are being pushed as far as possible into the net. These effects are exacerbated in an environment where the fact database is changing rapidly. Here the inflow of new facts create a large number of mostly unimportant rule activations and deactivations, when relevant facts are removed in favor of more recent data. Limiting the Rete net activity to those rules that are appropriate to the current focus of attention brings significant savings in avoiding unneeded matching. Human experts often employ a combination of forward and backward reasoning (Geogeff and Bonollo 1983). A control mechanism that consists solely of forward reasoning does not effectively model this process. In response to this, in many cases expert system programmers have built a goal-directed structure into their programs by adding a clause at the beginning of each rule to select the context of the goal for which the rule applies. These clauses then match facts that the programmer asserts manually to invoke the goal. This reduces the set of applicable rules and effectively provides backward reasoning. However, this approach is clumsy, and adds extra overhead for the programmer, who has to perform housekeeping to ensure the timely removal of goal facts, and settle auxiliary issues like conflicts between two concurrent goals. Also, this practice is only a convention, not supported or checked by the expert system language. We saw these problems arising from the attempt to use data to fulfil the role of control. What is really needed is a new control structure, that manages the goals cleanly and properly with a minimum of effort. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://ftp.cs.ucla.edu/pub/homeier/papers/eclips.pdf |
| Alternate Webpage(s) | http://www.trustworthytools.com/publications/eclips.pdf |
| Alternate Webpage(s) | http://trustworthytools.com/publications/eclips.pdf |
| Alternate Webpage(s) | http://www.cis.upenn.edu/~homeier/publications/eclips.ps |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
