AUDITOR'S ASSISTANT: A knowledge engineering tool for audit decisions;

              4
AUDITOR'S ASSISTANT: A Knowledge Engineering Tool For Audit Decisions*
Glenn Shafer, Prakash P. Shenoy, Rajendra P. Srivastava
University of Kansas
1. Introduction
In recent years, there has been significant interest in developing expert systems for assistance in audit decisions [see e.g, Boritz and Wensley, 1988; Chandler, 1985; Hansen and Messier, 1986a, and 1986b; Leslie et al., 1986]. It is believed that use of such systems will facilitate audit decisions and make audits more efficient and effective. This appears to be the reason that major accounting firms are committing increasingly greater resources to developing such systems [see e.g., Boritz and Brown, 1986; Kelly, 1987; Shpilberg and Graham, 1986].
Most of the expert systems being developed are rule-based. While such systems have many attractive features such as modularity of knowledge-base, ease of updating knowledge-base, etc., they are not well-suited for coherent reasoning under uncertainty. This is because in rule-based systems, the user has no control over the chain of inference whereas, coherent reasoning under uncertainty requires controlled firing of rules [Shafer, 1987]. Because of this difficulty, some developers of expert systems have avoided dealing with uncertainties altogether [Kelly et al., 1986]. In domains where uncertain reasoning is unavoidable, heuristic approaches have been attempted with little success [Shortliffe and Buchanan, 1975; Duda et al., 1976]. In recent years, considerable theoretical work has been done on the subject of coherent uncertain inference using Bayesian probabilities and belief-functions [see e.g., Pearl, 1986; Kong, 1986; Shenoy and Shafer, 1986; Mellouli, 1987; Shafer, Shenoy and Mellouli, 1987; Lauritzen and Spiegelhalter, 1988; Shafer and Shenoy, 1988]. The expert system described in this article represents one of the first practical applications of these new techniques.
The purpose of this paper is to describe an interactive tool called AUDITOR'S ASSISTANT (AA). The system, when fully developed, should
* This research has been supported in part by grants from the Peat Marwick Foundation, the
National Science Foundation grant No. IST-8610293 and General Research Fund of the University
of Kansas. The authors are grateful for discussions and assistance with programming from Yen-Teh
Hsia, Debra Zarley and Ragu Srinivasan.
61