Support and Plausibility Degrees in Generalized Functional Models

Paul-André MonneyUAI 1997

TL;DR

This paper introduces "generalized functional models" as a flexible framework for reasoning under uncertainty, particularly when traditional probabilistic models fall short due to a lack of prior information. It leverages concepts from Fisher's fiducial probability to define "support" and "plausibility" functions, offering a more nuanced way to express degrees of belief than standard probability distributions.

What Problem Does This Paper Address?

Many real-world scenarios involve inferring an unknown state (e.g., pregnancy, a policy choice) based on uncertain observations. Traditional Bayesian approaches require precise prior probabilities, which are often unavailable or difficult to elicit. This paper addresses the challenge of quantifying uncertainty and making inferences in situations where such priors are missing or where the relationship between observations and the underlying state is complex. It aims to provide a more natural and less assumption-laden framework for these types of problems, drawing inspiration from R.A. Fisher's work on fiducial probability.

Methods

The core contribution is the development of generalized functional models. These models are defined by a function $f$ that describes the interaction between a parameter variable ($\theta$), an observation variable ($\xi$), and a random source ($\omega$), along with a probability distribution $P$ for the random source $\omega$. Unlike traditional models, generalized functional models do not require a unique mapping from observations to parameters.

The analysis of these models leads to the concepts of support and plausibility functions. These functions are derived from a "hint" object, which encapsulates the knowledge gained from observations. The support function quantifies the degree to which an observation definitely supports a hypothesis, while the plausibility function quantifies the degree to which an observation is compatible with a hypothesis. These functions are shown to be related to belief functions in Dempster-Shafer theory. The paper also introduces a method for combining hints from multiple observations, analogous to Dempster's rule of combination.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and uses illustrative examples (Jessica's pregnancy test, policy identification) rather than empirical evaluation on datasets.

Key Insights and Ideas

  • Flexibility beyond Priors: Generalized functional models offer a way to reason under uncertainty even when precise prior probabilities are unavailable, a significant advantage over strict Bayesian approaches.
  • Support vs. Plausibility: The distinction between support (degree of definite evidence) and plausibility (degree of compatibility) provides a richer representation of uncertainty than a single probability value. This aligns with how humans often reason about evidence.
  • Fisher's Legacy: The framework elegantly extends ideas from Fisher's fiducial probability, providing a modern, formal interpretation for some of his reasoning techniques.
  • Hint Combination: The proposed method for combining hints from multiple observations offers a principled way to update knowledge as more data becomes available, mirroring Dempster's rule of combination.
  • Model Granularity: Generalized functional models can distinguish between situations that might appear similar under simpler "distribution models," suggesting they offer a more refined way to capture underlying real-world dynamics.

Key Results

The paper demonstrates how to construct support and plausibility functions for various scenarios, including:

  • Jessica's Pregnancy Test: With a test reliability of 0.9, the support for Jessica being pregnant is 0.1, and for not being pregnant is 0.9. This aligns with intuition and the test's reliability.
  • Policy Identification: The framework can distinguish between different policies based on observed outcomes, providing degrees of support and plausibility for each policy. For instance, if a "tails" outcome is observed in a coin-flipping scenario, the support for Peter using policy 1 becomes 1.

The paper also shows that the combination of hints from multiple observations can lead to either a precise hint (resulting in a probability measure) or a non-precise hint, depending on the observations.

Limitations and Future Work

  • Computational Complexity: While the theory is developed, the paper hints at the need for "special computational procedures involving the so-called commonality functions" for practical application, suggesting that direct computation of support and plausibility degrees might be complex.
  • Real-world Validation: The paper primarily uses illustrative examples. Further work would involve applying these models to more complex, real-world problems and comparing their performance against existing methods.
  • Interpretation of "Random Source": The role and nature of the "random source" ($\omega$) and its distribution $P$ are central. Further exploration into how to best define and elicit these components in practice would be valuable.
  • Connection to Other Theories: While connections to Dempster-Shafer theory are made, further exploration of the precise relationships and potential synergies with other evidence-based reasoning frameworks could be beneficial.

Summary updated

BibTeX
@inproceedings{monney1997support,
  author = {Paul-André Monney},
  title = {{Support and Plausibility Degrees in Generalized Functional Models}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1997},
}
Figure