A Sensitivity Analysis of Pathfinder: A Follow-Up Study

Keung-Chi Ng, Bruce AbramsonUAI 1991

TL;DR

This paper revisits a surprising finding from a previous study where a diagnostic AI system (Pathfinder) showed extreme sensitivity to its input parameters. The follow-up study refines the sensitivity analysis method to avoid altering absolute probabilities (0.0 and 1.0) and finds that Pathfinder does exhibit a "flat maximum" phenomenon, meaning small parameter changes have little impact on performance, aligning with expectations from decision analysis.

What Problem Does This Paper Address?

Artificial intelligence (AI) and decision analysis (DA) fields often intersect, particularly in how they handle uncertainty. While DA has well-established methods for analyzing how sensitive a decision is to changes in its input parameters (often showing a "flat maximum" where small changes have little effect), applying these to complex AI systems, which can have hundreds of variables and produce outputs beyond simple decisions, is less straightforward. A previous study on the Pathfinder diagnostic system surprisingly found it to be highly sensitive to parameter changes, contradicting the expected flat maximum. This paper addresses this discrepancy by conducting a follow-up sensitivity analysis with a refined methodology to better understand Pathfinder's behavior and its implications for extending DA sensitivity analysis to AI.

Methods

The core of this paper is a refined sensitivity analysis technique applied to the Pathfinder AI system. The previous study had a flaw: it varied probabilities that were exactly 0.0 or 1.0, which represent absolute certainty and shouldn't be perturbed. The new method addresses this by:

  • Conditional Probability Perturbation: Instead of directly adding noise to probabilities, they transform probabilities p in the range (0.0, 1.0) to a value v in the range (-∞, ∞) using the function v = ln(p / (1-p)). Noise is then added to v to create v', which is then transformed back to a probability p'. This ensures that probabilities close to 0.0 or 1.0 are less affected by noise than those closer to 0.5, and importantly, probabilities that are exactly 0.0 or 1.0 are left unchanged.
  • Two Sets of Priors: The analysis used both expert-assessed prior probabilities and a uniform distribution across all diseases to assess the impact of prior beliefs.
  • Multiple Noise Functions: Various noise functions (uniform and normal distributions with different standard deviations) were applied to the conditional probabilities.

Datasets

Not discussed in the paper. The study uses a body of 60 "classic" cases with known diagnoses for evaluation.

Key Insights and Ideas

  • Revisiting the "Flat Maximum": The most significant insight is that, with a more appropriate sensitivity analysis method, Pathfinder does exhibit the expected "flat maximum" phenomenon. This suggests the previous surprising results were an artifact of the analysis method, not an inherent property of the system.
  • The Importance of Endpoints: The paper highlights that probabilities at the extreme ends (0.0 and 1.0) represent absolute certainty and should not be perturbed in sensitivity analyses. Perturbing these values can lead to misleading conclusions about system robustness.
  • Extending DA to AI: The study demonstrates a viable way to extend traditional DA sensitivity analysis techniques to larger, more complex AI systems, providing a more nuanced understanding of their parameter robustness.

Key Results

The follow-up study, using the refined methodology, showed that:

  • Small parameter variations have little effect: Unlike the initial study, slight variations to Pathfinder's parameters now have minimal impact on its diagnostic performance.
  • Performance is more stable: The percentage of correct diagnoses and average confidence remained largely comparable to the original knowledge base, even with some added noise.
  • Larger refinements can still be meaningful: While small changes are inconsequential, larger, more systematic refinements to probabilities (especially those not at the endpoints) can still lead to improvements in performance, as indicated by the "Percentage Better" column in the tables.

Limitations and Future Work

  • Limitations: The authors acknowledge that their previous study's results were misleading due to the flawed methodology. The current study corrects this but also notes that the "flat maximum" observed is for small parameter changes. Larger, systematic changes might still reveal interesting behaviors.
  • Future Work: The paper suggests that this refined sensitivity analysis approach can be a valuable tool for understanding and validating complex AI systems, potentially leading to more robust and reliable AI models. The authors also emphasize the importance of parameters, especially those close to 0.0 and 1.0, for belief networks.

Summary updated

BibTeX
@inproceedings{ng1991sensitivity,
  author = {Keung-Chi Ng and Bruce Abramson},
  title = {{A Sensitivity Analysis of Pathfinder: A Follow-Up Study}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1991},
}
Figure