Focusing on the Most Important Explanations: Decision-Theoretic Horn Abduction

Paul O'RorkeAAAI 1994

Abstract

In safety-critical domains, identifying the most probable cause of a failure is often less critical than identifying the most consequential one. This paper introduces Decision-Theoretic Horn Abduction (DTHA), a method that prioritizes explanations by their "importance"—a metric combining Bayesian probability with decision-theoretic utility. By employing a best-first search strategy, DTHA focuses on high-stakes events, such as catastrophic failures, even when they are statistically less likely than benign issues. The approach is validated through a Failure Modes and Effects Analysis (FMEA) of an automotive fuel system model.

Introduction

Abductive reasoning—the process of inferring the most plausible causes for observed effects—is a cornerstone of diagnostic AI. Traditionally, research has centered on finding the "most likely" explanation. However, in fields like medicine or aerospace engineering, a likelihood-only criterion can be insufficient. A low-probability event with a catastrophic outcome (e.g., a rare but fatal engine fire) often demands more immediate attention than a high-probability event with a trivial outcome (e.g., a common sensor glitch). DTHA provides a formal mechanism to focus an AI’s attention on these high-stakes explanations without getting bogged down in an infinite sea of possible causes.

Methodology

DTHA extends David Poole’s Probabilistic Horn Abduction (PHA) by incorporating utility. The algorithm calculates the Importance of an explanation as the product of its probability ($p$) and its utility (cost).

The system utilizes a best-first search algorithm to maintain a set of partial explanations. At each iteration, it expands the path with the highest current importance. This leads to dynamic focus shifting: as the algorithm works backward and makes more assumptions, the probability of a specific path typically decreases. If the importance of the current path falls below that of a competing explanation for a different outcome, the algorithm shifts its focus to the more critical candidate.

Case Study: Automotive FMEA

The authors demonstrate DTHA using a custom model of an automotive system, illustrated in Figure 1. The model includes components such as a fuel pump, delivery pipes, an injection system, and a combustion chamber.

Schematic of the automotive system model used for DTHA evaluation.

Failure modes like broken pipes or leaky joints are assigned prior probabilities, while outcomes like engine fire (F), high emissions (H), and low efficiency (L) are assigned costs (negative utilities) of $10^6$, $10^5$, and $10^4$ respectively.

Results and Key Insights

The DTHA algorithm successfully prioritized high-cost failures over high-probability ones. For instance, the system prioritized a pipe joint leak (weighted cost $10^4$) over a sensor failure (weighted cost $5 \times 10^3$), despite the sensor failure being five times more probable.

Table of causes, consequences, and their resulting priorities in the DTHA framework.

As shown in the priority table above, the system maintains a clear hierarchy of investigation based on the severity of the consequence. The algorithm also demonstrated high efficiency by "jumping" between different outcomes as the search progressed. This backward-chaining approach allows for the discovery of unanticipated failure sequences that traditional forward-chaining FMEA tools might miss.

Discussion and Limitations

While DTHA offers a robust framework for prioritized diagnosis, it relies on the availability of precise utility and probability values, which can be difficult to quantify in practice. Furthermore, the current model assumes component independence and does not yet incorporate "detectability"—a key component of standard Risk Priority Numbers (RPN) in FMEA. Future work will explore more sophisticated utility calculations for complex, interdependent systems where one failure may trigger a cascade of physical stresses.

Summary updated

BibTeX
@inproceedings{ororke1994focusing,
  author = {Paul O'Rorke},
  title = {{Focusing on the Most Important Explanations: Decision-Theoretic Horn Abduction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure