TL;DR
This paper introduces a novel probabilistic framework for handling uncertainty by allowing for "nonmeasurable" events, which are events for which a precise probability cannot be assigned. Instead, belief in such events is represented by an interval of inner and outer measures, bridging the gap between classical probability and Dempster-Shafer theory of evidence. This approach offers a more nuanced way to reason about uncertainty and reveals a deep connection between these two seemingly distinct formalisms.
What Problem Does This Paper Address?
The dominant approach to dealing with uncertainty in AI is Bayesian probability, which requires assigning a precise numerical probability to every event. While theoretically sound, this can be computationally challenging and may not always reflect human intuition, as people are often comfortable with probability ranges rather than exact values. Critics argue that forcing precise probabilities can be unrealistic, especially when information is incomplete. Alternative approaches like Dempster-Shafer theory offer intervals for belief but are sometimes viewed as less grounded in classical probability. This paper aims to reconcile these perspectives by developing a probabilistic framework that naturally accommodates uncertainty intervals and demonstrates their equivalence to key aspects of Dempster-Shafer theory.
Methods
The core innovation is the introduction of "nonmeasurable events" into a probabilistic framework. Traditionally, nonmeasurable sets arise from mathematical necessity in probability theory. Here, they are embraced by choice to represent events where an agent lacks sufficient information for a precise probability assignment. For these nonmeasurable events, the paper uses the concepts of inner measure and outer measure, which are derived from a standard probability measure. The inner measure represents the probability of the largest measurable event contained within the nonmeasurable event, while the outer measure represents the probability of the smallest measurable event containing it. This effectively defines an interval [inner measure, outer measure] for the belief in the event. The paper then proves that these inner measures induced by probability measures are mathematically equivalent to Dempster-Shafer belief functions.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Embracing Nonmeasurability: The paper reframes mathematical "nuisance" (nonmeasurable sets) as a powerful tool for representing incomplete knowledge and uncertainty.
- Interval-Based Belief: Instead of a single probability, belief in an event is represented by an interval, offering a more flexible and intuitive representation of uncertainty.
- Unifying Probability and Dempster-Shafer Theory: A key finding is that inner measures induced by probability measures are equivalent to Dempster-Shafer belief functions. This provides a strong theoretical foundation for Dempster-Shafer theory within classical probability.
- Beyond Additivity: The framework naturally handles situations where the belief in an event and its negation do not sum to 1, a common characteristic of non-probabilistic reasoning.
Key Results
The paper establishes a precise equivalence between:
- Inner measures induced by probability measures and Dempster-Shafer belief functions.
- Outer measures induced by probability measures and Dempster-Shafer plausibility functions.
This equivalence is demonstrated through theorems that show a natural correspondence between Nilsson structures (representing precise probabilities for propositional formulas) and measurable probability structures, and between probability structures and Dempster-Shafer structures. The paper also presents an axiomatization (AX MEAS and AX) for reasoning about these probabilistic and belief structures, showing that the complexity of deciding satisfiability for these formulas is NP-complete, matching that of propositional logic.
Limitations and Future Work
The authors acknowledge that their approach is a theoretical framework and that practical implementation, especially for complex systems, would require further development. They also mention that their work on combining evidence is a natural extension of Dempster's rule of combination and that they are investigating alternatives, particularly focusing on conditional probability with respect to nonmeasurable sets. Future work includes reporting on this research in a subsequent paper.
Summary updated