Some extensions of probabilistic logic

Su-Shing ChenUAI 1986

TL;DR

This paper extends Nilsson's probabilistic logic by incorporating the Dempster-Shafer theory of evidence. This allows for reasoning with uncertain and incomplete information, moving beyond the limitations of traditional probabilistic approaches by handling logically inconsistent interpretations and providing a more nuanced representation of belief.

What Problem Does This Paper Address?

Traditional probabilistic logic, as proposed by Nilsson, assigns truth values to propositions as probabilities between 0 and 1. While this approach reduces to classical logic when probabilities are 0 or 1, it inherits limitations from Bayesian reasoning, particularly in handling situations where propositions might be logically inconsistent or where belief is not precisely quantifiable. The paper aims to address these limitations by developing an "evidential logic" that can better represent and reason with uncertainty, especially in AI applications where assumptions of consistency might be unnatural.

Methods

The core of the paper is the extension of probabilistic logic into an "evidential logic" using the framework of Dempster-Shafer theory. This involves:

  • Frames of Discernment: Considering the space of all possible interpretations (truth assignments) for a set of logical propositions, including logically inconsistent ones.
  • Basic Probability Assignment (bpa): Defining a function m that assigns a numerical value (belief mass) to each subset of interpretations, representing the degree of belief in that subset. This generalizes traditional probability distributions.
  • Belief and Plausibility Functions: Deriving belief (Spl) and plausibility (Pls) functions from the bpa, which provide lower and upper bounds on the probability of a proposition. These are represented as evidential intervals [Spl(s), Pls(s)].
  • Dempster's Combination Rule: Utilizing this rule to combine evidence from different sources, allowing for the updating of beliefs.
  • Semantic Tree Representation: Employing a binary semantic tree to represent interpretations and facilitate the assignment of truth values and the computation of probabilities.
  • Semantics as Random Variables: Formalizing the probabilistic logic within the framework of multi-dimensional random variables, where each proposition can be seen as a random variable with outcomes 0 or 1. This allows for the computation of joint, marginal, and conditional probabilities.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Probabilistic Logic as a Subset of Probability Theory: The paper argues that Nilsson's probabilistic logic, when viewed through the lens of multi-dimensional random variables and joint probabilities, is fundamentally subsumed by standard probability theory.
  • Evidential Logic for Inconsistent Interpretations: The key innovation is the extension to evidential logic, which explicitly handles logically inconsistent interpretations, a common scenario in real-world AI problems.
  • Evidential Intervals for Belief: Representing belief in propositions as intervals [Spl(s), Pls(s)] provides a more flexible and informative way to express uncertainty than single probability values, especially when dealing with incomplete or conflicting evidence.
  • Dempster-Shafer for AI Reasoning: The paper highlights the applicability of Dempster-Shafer theory and its combination rule for evidential reasoning in AI, suggesting it offers advantages over purely Bayesian approaches in certain contexts.
  • Connection to Relaxation Labeling: The proposed evidential logic provides a framework for relaxation schemes, which are useful in areas like computer vision for assigning labels (truth values) to propositions iteratively until a stable state is reached.

Key Results

  • The paper demonstrates how to define evidential intervals for logical propositions using basic probability assignments derived from frames of discernment.
  • It outlines an extension of Nilsson's probabilistic inference scheme that incorporates the Dempster-Shafer framework, allowing for the calculation of evidential intervals for new propositions.
  • The authors show that the probabilistic interpretation of probabilistic logic can be viewed as a relaxation scheme that converges to a stable labeling of logical propositions.

Limitations and Future Work

  • Generalization of Nilsson's Scheme: The paper presents an extension of Nilsson's probabilistic inference scheme, suggesting that it can be applied to new propositions by constructing an extended semantic tree and finding a new bpa.
  • Hierarchical Hypothesis Spaces: The authors note that the binary semantic tree is well-suited for dealing with hierarchical hypothesis spaces, implying potential for future work in structured reasoning.
  • Convergence of Relaxation: The paper suggests that the process of evidential reasoning, viewed as a relaxation scheme, will converge to a stable and deterministic labeling of logical propositions, indicating a direction for further analysis of convergence properties.

Summary updated

BibTeX
@inproceedings{chen1986some,
  author = {Su-Shing Chen},
  title = {{Some extensions of probabilistic logic}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1986},
}
Figure