Revising Imprecise Probabilistic Beliefs in the Framework of Probabilistic Logic Programming

Anbu Yue, Weiru LiuAAAI 2008

Abstract

This paper addresses the integration of belief revision and probabilistic reasoning by proposing a framework for revising imprecise probabilistic beliefs in Probabilistic Logic Programming (PLP). We introduce a set of eight postulates (R1-R8) that extend the AGM and Darwiche-Pearl frameworks to handle probability intervals and conflicting evidence. The core of our approach is a revision operator based on minimal change principles, which we prove satisfies a representation theorem. This framework generalizes classical Bayesian conditioning and Jeffrey’s Rule while providing a robust mechanism for iterated revision in the presence of imprecision.

Introduction

In classical artificial intelligence, the AGM framework for belief revision governs how an agent should update its knowledge when faced with information that contradicts its current beliefs. Conversely, probabilistic reasoning typically relies on Bayesian conditioning. However, these two paradigms often clash. Bayesian methods struggle when new evidence is assigned a prior probability of zero, and they typically assume precise probability distributions. In real-world scenarios, knowledge is often imprecise—represented by intervals rather than single values—and agents must be able to revise their beliefs even when surprised by "impossible" events.

This paper bridges this gap by introducing a formal framework for revising imprecise probabilistic beliefs within PLPs. By treating beliefs as sets of probability distributions, the authors provide a method for incorporating new conditional probabilistic formulas while adhering to the principle of minimal change.

Theoretical Framework and Postulates

The authors adopt an axiomatic approach, defining the characteristics of a "rational" revision operator through eight postulates (R1-R8). These postulates are heavily influenced by the Darwiche-Pearl (DP) postulates for iterated belief revision but are adapted for the probabilistic context. They ensure that the revision is sensitive to the bounds of the evidence and follows the principle of minimal change.

A key concept introduced is perpendicularity, which formalizes the notion of irrelevance. If new information pertains to a domain of events that is "perpendicular" to existing beliefs, those beliefs should remain unchanged. This ensures that the revision process is local and does not cause unnecessary global shifts in the agent's epistemic state.

The Revision Operator and Minimal Change

The proposed revision operator ($\bullet$) is designed to select probability distributions from the original set $\Psi_P$ that are "closest" to the new evidence $\mu$. This closeness is measured using a quasi-distance $d_\phi$. Geometrically, if we view the set of satisfying distributions as a convex set, the revision process involves finding the point in the new evidence space that minimizes the deviation from the current set.

Table 1: Possible worlds and probability distributions for the 'Birds, Penguins, and Flying' scenario.

As shown in Table 1, the framework operates over a set of possible worlds (e.g., $I_1$ to $I_8$ in the classic "Birds, Penguins, and Flying" scenario). Each world represents a unique combination of predicates, and the revision operator adjusts the probability mass ($x_i, \epsilon_i$) assigned to these worlds to satisfy new constraints ($x'_i, \epsilon'_i$). This specific example is used throughout the paper to demonstrate how the system handles conflicting conditional information, such as learning that a specific bird is a penguin and therefore likely cannot fly.

Key Results and Insights

The mathematical foundation of the paper is the Representation Theorem (Theorem 1). This theorem provides the necessary and sufficient conditions for a revision operator to be considered "rational" under their framework, linking the axiomatic postulates to the mathematical operator.

Furthermore, the framework demonstrates several significant properties:

  • Generalization of Classical Rules: The operator collapses into Jeffrey’s Rule or Bayesian conditioning when beliefs are precise. If the evidence is a certain fact (probability 1), it aligns with the DP postulates.
  • Iterated Revision: The framework successfully handles multiple sequential updates ($(\Psi \ast \mu_1) \ast \mu_2$) without losing consistency, satisfying Postulate R6 and ensuring the agent does not over-react to minor changes in evidence.
  • Probabilistic Ramsey Test: The authors adapt the famous Ramsey Test to the probabilistic setting, showing that a conditional is accepted if and only if revising by the antecedent leads to the consequent.

Conclusion and Future Work

While the framework provides a robust theoretical basis for revising imprecise beliefs, several challenges remain. The current postulates focus on single probabilistic formulas; revising by a conjunction of formulas is significantly more complex and remains beyond the scope of this work. Additionally, the computational cost of optimizing over convex sets of distributions in large-scale PLPs needs further investigation. Finally, the method avoids "distributions of distributions," which simplifies the logic but may limit the representation of meta-uncertainty. Future research will focus on these computational aspects and extending the framework to more complex evidence structures.

Summary updated

BibTeX
@inproceedings{yue2008revising,
  author = {Anbu Yue and Weiru Liu},
  title = {{Revising Imprecise Probabilistic Beliefs in the Framework of Probabilistic Logic Programming}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure