Diagnostic expert systems constructed using traditional knowledge-engineering techniques identify malfunctioning components using rules that associate symptoms with diagnoses. Model-based diagnosis (MBD) systems use models of devices to find faults given observations of abnormal behavior. These approaches to diagnosis are complementary. We consider hybrid diagnosis systems that include both associational and model-based diagnostic components. We present results on explanation-based learning (EBL) methods aimed at improving the performance of hybrid diagnostic problem solvers. We describe two architectures called EBL_IA and EBL(p). EBL_IA is a form fo "learning in advance" that pre-compiles models into associations. At run-time the diagnostic system is purely associational. In EBL(p), the run-time diagnosis system contains associational, MBD, and EBL components. Learned associational rules are preferred but when they are incomplete they may produce too many incorrect diagnoses. When errors cause performance to dip below a give threshold p, EBL(p) activates MBD and explanation-based "learning while doing". We present results of empirical studies comparing MBD without learning versus EBL_IA and EBL(p). The main conclusions are as follows. EBL_IA is superior when it is feasible but it is not feasible for large devices. EBL(p) can speed-up MBD and scale-up to larger devices in situations where perfect accuracy is not required.

An important aspect of intelligent behavior is the ability to reason, make decisions, and act in spite of uncertainty. This paper presents a qualitative logic of decision that supports decision-making under uncertainty. To be specific, the paper presents a knowledge representation language based upon subjective Bayesian decision theory that aims to capture some aspects of common-sense reasoning associated with making decisions about actions. The language addresses the problem of describing justifications of rational choices in situations where the alternatives involve trading off potential losses and gains. The logic and an associated qualitative arithmetic are implemented in an efficient PROLOG program. Examples illustrate their use in several concrete decision-making situations.

In this paper we study empirically the behavior of algorithm structure-based abduction (SAB) which was developed in the framework of constraint networks [5], by comparing it with a model-based diagnosis algorithm MBD, simialr to algorithm GDE [3]. The distinguishing features of algorithm SAB are that it exploits the structure of the constraint network and that it is most efficient when the problem contains no cycles. The performance of both algorithms is tested on a family of parametrized acyclic combinational circuits for the task of finding all minmal-cardinality diagnoses. The results show that due to its exponential complexity for large circuits MBD can run out of space and time, while SAB is able to compute the diagnoses for those same circuits. Unlike MBD, SAB appears to be insensitive to variations in the circuit types and input probabilities. For small circuits, the average time for MBD relative to SAB seems to be proportional to the number of conflicts relative to the number of diagnoses.

This paper describes two methods for integrating model-based diagnosis (MBD) and explanation-based learning. The first method (EBL) uses a generate-test-debug paradigm, generating diagnostic hypotheses using learned associational rules that summarize model-based diagnostic experiences. This strategy is a form of "learning while doing" model-based troubleshooting and could be called "online learning." The second diagnosis and learning method described here (EEL-STATIC) involves ''learning in advance." Learning begins in a training phase prior to performance or testing. Empirical results of computational experiments comparing the learning methods with MBD on two devices (the polybox and the binary full adder) are reported. For the same diagnostic performance, EBL-STATIC is several orders of magnitude faster than MBD while EBL can cause performance slow-down.

Diagnostic expert systems constructed using traditional knowledge-engineering techniques identify malfunctioning components using rules that associate symptoms with diagnoses. Model-based diagnosis (MBD) systems use models of devices to find faults given observations of abnormal behavior. These approaches to diagnosis are complementary. We consider hybrid diagnosis systems that include both associational and model-based diagnostic components. We present results on explanation-based learning (EBL) methods aimed at improving the performance of hybrid diagnostic problem solvers. We describe two architectures called STATIC and EBL(p). STATIC pre-compiles models into associations, and at run-time the diagnostic system is purely associational. In EBL(p), the run-time diagnosis system is a hybrid: learned associational rules are preferred but the MBD component is activated whenever the performance falls below a threshold p. We present results of empirical studies comparing MBD without learning versus STATIC and EBL(p). The main conclusions are as follows. STATIC is superior when it is feasible but it is not feasible for large devices. EBL(p) can speed-up MBD and scale-up to larger devices in situations where perfect accuracy is not required.

Model-based diagnosis (MBD) provides several advantages over experiential rule-based systems. A principal shortcoming of MBD is that MBD learns nothing from any given example. An MBD system facing the same task a second time will incur the same computational effort as that incurred the first time. Our earlier work on incorporating explanation-based learning (EBL) in MBD [4] suggested a diagnostic architecture integrating EBL and MBD components. In this architecture, EBL was used to learn diagnostic rules. But the diagnoses proposed by the rules could be erroneous. So constraint suspension testing was used to check all proposed diagnoses. Insisting on perfect accuracy causes the performance of this scheme for "learning while doing" to deteriorate rapidly with the size of the device to be diagnosed. In this paper, we describe a method for trading off accuracy for efficiency. In this approach, most diagnosis problems are handled by the associational rules learned from previous problems. Model-based reasoning and learning are activated only when performance drops below a given threshold. We present empirical results on circuits of increasing number of components illustrating how this approach scales up.