There is an ongoing debate as to whether model-based or case-based diagnostic expert systems are superior. Our experience has shown that the two are not mutually exclusive and, to the contrary, complement each other. Current expert system technology is capable of two reasoning mechanisms, in addition to other mechanisms, integrated into one system. Depending on the knowledge available, and time and cost considerations, expert systems allow the user to decide the relative proportion of case-based to model-based reasoning to employ in any given situation. Diagnostic support software should be evaluated by two critical factor groups: (a) cost and time to deployment, and (b) accuracy, completeness and efficiency of the diagnostic process. In this paper we will discuss the role of expert systems in combining model-based and case-based reasoning to effect the most efficient user defined solution to diagnostic performance.

This paper describes our experience in using a diagnostic expert system (AITEST) for fault isolation, diagnosis and repair action recommendations for a remote, tele-controlled UUT (Unit Under Test). We identify the unique requirements of diagnosis in this class of problems, and explains how these requirements were met in less than three person-months by integrating the expert system with specially-developed tele-control software and test software to form a seamless and capable diagnostic system for a specific UUT. We describe the process we used in analyzing the problem, concurrent development of knowledge base and test software, and their integration into the complete solution. We conclude by presenting preliminary results.>

A technique for recognizing a 2-D unoccluded polygonal object by combining the alignment method with efficient string matching algorithms is presented. The approach is based on a single anchor point: the gravitation center of the contour (GCC) of the object. The GCC is stable and insensitive to digitization errors, and it can always be found and calculated very efficiently. Additionally, it is universal and does not depend on the specific library of known objects. In this approach, most of the work is done in the preprocessing stage, when all of the library objects are transformed into a GCC canonical representation form. For the recognition stage, the canonical representation of the unknown object concatenated to itself is considered as a 'text' and the canonical representation of the known library objects as a set of 'words'. The recognition problem is then reduced to finding all of the occurrences of the 'words' in the 'text', for which an efficient O(n) algorithm is introduced, where n is the order of the polygon being recognized. This approach was evaluated with several sets of objects from different fields, and very satisfactory results were obtained.>

In order to establish a uniform basis for comparing and evaluating diagnostic expert systems for functional electronic troubleshooting, a real-life benchmark unit is proposed. The HP-3478P unit is essentially identical to the HP-3478A digital multimeter. The authors first outline the criteria that guided them in the selection of this unit from a wide variety of real-life units. They describe the structure of the HP-3478P and its test set. They then provide an illustrative test case and a discussion of the complexity of the unit. The HP-3478P was used to evaluate the AITEST system on a wide variety of cases, and it was found to be very useful.>

Contextual template matching refers to situations where the feature sets of the two templates under consideration may not be identical, but are rather contingent on the context within which these templates are observed. This paper addresses context-dependent template matching where the relevancy of certain features to a given object is contingent on the value of higher level features. In such situations portions of the patterns for two objects X and Y may not be comparable, and therefore classical distance measures are not applicable. A framework for describing pattern spaces with hierarchically dependent features-named conditional pattern spaces (CPS)-is presented, and a distance measure for comparing objects in a CPS is developed and illustrated on archaeological objects. This measure satisfies several desired properties of distance measures, such as the metric requirements.

Information acquisition strategies which incorporate human heuristics are formulated for pattern classification tasks, and their effectiveness is evaluated. The heuristics are based on two observations: (1) human decision-makers tend to limit themselves to a subset of classes and to select features oriented toward this subject only; (2) human decision-makers typically use considerations related to the history of process, that is, to class probabilities in earlier stages, while classical Bayesian strategies consider only the current class probabilities. These heuristics are incorporated in four different strategies with which the authors experimented. The findings are useful for the development of decision aids whose information selection strategies may be tuned to the operator's information selection behaviour by offering the operator an aid which reflects his own information priorities.

Performance evaluation measures for multimembership classifiers are presented and applied in a retrospective study on the diagnostic performance of the MEDAS (Medical Emergency Decision Assistance System) system. Admission and discharge diagnoses for 122 patients with one or more of 26 distinct disorders in five major disorder categories were gathered. The average number of disorders per patient was 2 with 36 (29.5 percent) patients having 3 or more disorders simultaneously. The features (symptoms, signs, and laboratory data) available at admission were entered into a multimembership Bayesian pattern recognition algorithm which permits for diagnosis of multiple disorders. When the top five computer-ranked diagnoses were considered, all of the correct diagnoses for 86.1 percent of the patients were displayed by the fifth position. In 71.6 percent of these cases, no false diagnosis preceded any correct diagnosis. In ten cases a discharge diagnosis which was suggested by the available findings was omitted by the admitting physician. In six of these ten cases, the overlooked diagnoses appeared at the computer ranked list above all false diagnoses. Considering the urgency of diagnosis in the Emergency Department, the high uncertainty involved due to the limited availability of data, and the high frequency with which multiple disorders coexist, this limited study encourages our confidence in the MEDAS knowledge base and algorithm as a useful diagnostic support tool.

Situation assessment tasks, e.g., medical diagnosis, battlefield reading, corporation status assessment for merger or acquisition purposes, are formulated as a general family of problem solving tasks. The generic nature of this family task as a multimembership hierarchical pattern recognition problem is characterized and the types of decision support systems (DSS) are identified. The focus is on knowledge representation and elicitation, although issues related to inference mechanisms, system structure, and expert-machine-user interface are also discussed. Two types of knowledge are distinguished: global knowledge and local knowledge. Global knowledge is required to determine directions on which to focus attention, while local knowledge is required for assessing the validity of a specific alternative based on a given set of findings. Global knowledge is represented as a network of relevancy pointers between alternatives and features. Attached to the links of this network are weights by which the strength of relevancy is evaluated and global directions (hypotheses) for situation analysis are determined. For local knowledge, it seems that in most practical problems multiple representation techniques would be required to characterize adequately the alternatives by means of their relevant features. The presentation is accompanied by examples for military situation assessment. However, comparable examples from medical and business applications are also cited. In fact, many of the ideas presented here have already been implemented in the MEDAS system-a medical DSS for emergency and critical care medicine.

The low sensitivity of the probability of error rule (Pe rule) for feature selection is demonstrated and discussed. It is shown that under certain conditions features with significantly different discrimination power are considered as equivalent by the Pe rule. The main reason for this phenomenon lies in the fact that, directly, the Pe rule depends only on the most probable class and that, under the stated condition, the prior most probable class remains the posterior most probable class regardless of the result for the observed feature. A rule for breaking ties is suggested to refine the feature ordering induced by the Pe rule. By this tie-breaking rule, when two features have the same value for the expected probability of error, the feature with the higher variance for the probability of error is preferred.

A knowledge-based interactive sequential diagnostic system is introduced which provides for diagnosis of multiple disorders in several body systems. The knowledge base consists of disorder patterns in a hierarchical structure that constitute the background medical information required for diagnosis in the domain under consideration (emergency and critical care medicine, in our case). Utilizing this knowledge base, the diagnostic process is driven by a multimembership classification algorithm for diagnostic assessment as well as for information acquisition [1]. A key characteristic of the system is congenial man-machine interface which comes to expression in, for instance, the flexibility it offers to the user in controlling its operation. At any stage of the diagnostic process the user may decide on an operation strategy that varies from full user control, through mixed initiative to full system control. Likewise, the system is capable of explaining to the user the reasoning process for its decisions. The model is independent of the knowledge base, thereby permitting continuous update of the knowledge base, as well as expansions to include disorders from other disciplines. The information structure lends itself to compact storage and provides for efflcient computation. Presently, the system contains 53 high-level disorders which are diagnosed by means of 587 medical findings.

The sensitivity of Bayesian pattern recognition models to multiplicative deviations in the prior and conditional probabilities is investigated for the two-class case. Explicit formulas are obtained for the factor K by which the computed posterior probabilities should be divided in order to eliminate the deviation effect. Numerical results for the case of binary features indicate that the Bayesian model tolerates large deviations in the prior and conditional probabilities. In fact, the a priori ratio and the likelihood ratio may deviate within a range of 65-135 percent and still produce posterior probabilities in accurate proximity of at most ±0.10. The main implication is that Bayesian systems which are based on limited data or subjective probabilities are expected to have a high percentage of correct classification despite the fact that the prior and conditional probabilities they use may deviate rather significantly from the true values.

The concept of irrelevant features in Bayesian models for pattern recognition is introduced, and its mathematical meaning is explained. A technique for computing the conditional probabilities of irrelevant features, if necessary, is described. The effect of irrelevant features on feature selection in sequential classification is discussed and illustrated.

Several rules for feature selection in myopic policy are examined for solving the sequential finite classification problem with conditionally independent binary features. The main finding is that no rule is consistently superior to the others. Likewise no specific strategy for the alternating of rules seems to be significantly more efficient.

The ranking of features according to a feature selection criteria is examined. The concept of\epsilon-equivalence is introduced to measure the extent to which a ranking deviates from the ranking included by the probability of error role. The relationship between the\epsilon-equivalence of a given role and the bounds on the probability of error derived from this rule is demonstrated. Illustrations of the\epsilon-equivalence concept are presented for Shannon's equivocation rule, the quadratic equivocation rule, and the Bhattacharyya rule. A numerical example concludes the presentation.

The basic properties of Renyi's entropy are reviewed, and its concavity properties are characterized. New bounds (referred to asI_{\alpha}bounds) on the probability of error are derived from Renyi's entropy and are compared with known bounds. It is proved that for the two-class case, theI_{2}bound is sharper than many of the previously known bounds. The difference between theI_{2}bound and the real value of the probability of error is at most 0.09.