On reasoning in networks with qualitative uncertainty

Simon Parsons, E. H. MamdaniUAI 1993

TL;DR

This paper introduces a novel method for qualitative reasoning under uncertainty that integrates probability, possibility, and belief theories. By analyzing the qualitative impact of evidence on network nodes, it offers a way to reason even when precise numerical data is unavailable and provides a framework for integrating different uncertainty formalisms.

What Problem Does This Paper Address?

In many real-world scenarios, precise numerical data for probabilistic reasoning is either unavailable or inappropriate. Traditional quantitative methods can struggle with this lack of precision, leading to knowledge bases that are only applicable in narrow domains. Qualitative approaches offer a higher level of abstraction, focusing on the direction of change (e.g., increase, decrease, no change) rather than exact values. While existing qualitative methods primarily focus on probabilistic reasoning, this paper addresses the need for a unified approach that can handle uncertainty expressed through probability, possibility theory, and Dempster-Shafer belief functions. The challenge lies in developing a method that can reason qualitatively across these different formalisms and integrate them effectively.

Methods

The core of the paper's contribution is a new approach to qualitative reasoning that leverages differential calculus on qualitative values. Instead of using numerical values, the method uses a restricted set of qualitative states: positive change [+], negative change [-], zero change [0], or unknown [?].

The key innovation lies in defining relationships for how the certainty value of one variable changes in response to changes in another. These relationships are:

  • Follows: The certainty value of X increases when Y's increases, and decreases when Y's decreases.
  • Varies inversely: The certainty value of X decreases when Y's increases, and increases when Y's decreases.
  • Independent: The certainty value of X does not change as Y's value changes.

These relationships are determined by analyzing the qualitative sign of the derivative of the certainty value of one variable with respect to another. The paper then applies this framework to simple network structures (single links and more complex networks with multiple parents) and derives qualitative propagation rules for probability, possibility, and belief theories. A crucial aspect is the use of a "monotonicity assumption" to bridge the gap between different formalisms: if a probability increases, its corresponding possibility and belief values do not decrease.

Datasets

Not discussed in the paper. The paper uses a hypothetical medical example network to illustrate its concepts.

Key Insights and Ideas

  • Unified Qualitative Reasoning: The paper demonstrates that a single qualitative framework can be applied to probability, possibility, and belief theories, unifying their reasoning capabilities.
  • Derivative-Based Qualitative Analysis: Using the sign of derivatives provides a principled way to determine qualitative relationships between variables, extending beyond simple "positive influence" definitions.
  • Bridging Formalisms: The monotonicity assumption offers a plausible and weak link for translating qualitative insights between different uncertainty formalisms, enabling integration.
  • Flexibility in Semantics: The approach is syntactically based, allowing different semantic interpretations to be attached to the qualitative values without invalidating the reasoning process.
  • Qualitative Integration: The method allows for reasoning about the impact of evidence expressed in one formalism on variables represented in another, facilitating a more holistic understanding of complex systems.

Key Results

The paper derives specific qualitative propagation rules for simple network links (A → C) and more complex networks (B&C → D) for probability, possibility, and belief theories. For instance, it shows how a change in p(a) affects p(c) based on conditional probabilities like p(c|a) and p(c|-a). It also demonstrates how to integrate these formalisms, as shown in the medical example where changes in probability (e.g., Sjorgen's syndrome) are used to infer changes in belief and possibility (e.g., patient's pain, vasculitic lesions). The results highlight that possibility theory can represent a wider range of behaviors than probability theory, while evidence theory is the least restricted.

Limitations and Future Work

  • Singly Connected Networks: The current analysis is primarily focused on singly connected networks. The paper acknowledges that handling multiply connected networks poses additional challenges.
  • Order of Magnitude Reasoning: A significant limitation of qualitative systems, including this one, is the inability to distinguish between small and large changes (e.g., 0.001 vs. 100,000). The authors suggest that future work could incorporate order of magnitude reasoning to address this.
  • Combination Rules: While the approach is flexible regarding combination rules, switching semantics might require re-deriving propagation conditions, as seen with Dempster's rule in Baldwin's voting model.
  • Automation: The authors express an intention to explore automating reasoning of the form "If A is bigger than B and B is bigger than C then A is bigger than C" in the near future.

Summary updated

BibTeX
@inproceedings{parsons1993reasoning,
  author = {Simon Parsons and E. H. Mamdani},
  title = {{On reasoning in networks with qualitative uncertainty}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}

Loading related content...

Figure