Belief Revision in Probability Theory

Pei WangUAI 1993

Abstract

In artificial intelligence and cognitive science, Bayesian probability is frequently championed as the optimal framework for reasoning under uncertainty. However, this paper argues that standard probability theory fails to distinguish between the propagation of logical consequences and the revision of beliefs based on new evidence. By analyzing the limitations of a single-value probability representation, author Pei Wang proposes the Non-Axiomatic Reasoning System (NARS), which utilizes frequency-confidence pairs to better model how an adaptive system merges conflicting information and tracks the weight of evidence.

Introduction

In the quest to build robust Artificial Intelligence, Bayesian probability has become the "gold standard" for handling uncertainty. It provides a rigorous mathematical framework for updating the likelihood of a hypothesis given new data. Yet, a fundamental question remains: is the calculation of a conditional probability the same thing as learning from experience?

Pei Wang argues that it is not. He identifies a critical gap in current AI methodologies: the conflation of propagation (deducing consequences within a fixed model) and revision (updating the model itself when new, potentially conflicting evidence arrives). While Bayes' theorem is an excellent tool for the former, it is often "insufficiently informative" for the latter.

The Conceptual Gap: Propagation vs. Revision

The core of Wang’s critique lies in the distinction between explicit and implicit conditions. In a standard Bayesian calculation, we determine $P(H|E)$—the probability of a hypothesis $H$ given explicit evidence $E$. However, this probability is always relative to a set of implicit background assumptions or context, which we can call $C$.

When we use Bayes' theorem, we are essentially performing propagation: we are asking what the probability of $H$ would be if $E$ were true, given our existing knowledge $C$. However, true revision occurs when $E$ actually becomes part of our knowledge base. Standard probability theory often treats this as a simple replacement (e.g., Jeffrey’s Rule), where new information overwrites the old. Wang argues that a truly adaptive system should instead merge the new evidence with the old, weighing each based on its respective reliability.

The "Single Number" Problem

A primary technical limitation of Bayesianism is its reliance on a single probability value. This creates a "blind spot" regarding the stability of a belief. Consider two scenarios:

  1. A coin is tossed twice, resulting in one head and one tail ($P=0.5$).
  2. A coin is tossed 1,000 times, resulting in 500 heads and 500 tails ($P=0.5$).

In a standard Bayesian system, both scenarios result in a probability of 0.5. However, the second scenario represents a much more "stable" belief. If the next toss is a head, the probability in the first scenario shifts dramatically, while in the second scenario, it barely moves. Because Bayes' theorem collapses the amount of evidence into a single number, the system loses track of how much data supports the current distribution, making it difficult to determine how much the belief should change when new data arrives.

The NARS Alternative: Frequency and Confidence

To address these limitations, Wang introduces the Non-Axiomatic Reasoning System (NARS). In NARS, the truth-value of a statement is not a single probability, but a pair of values:

  • Frequency ($f$): The proportion of positive evidence encountered so far.
  • Confidence ($c$): A measure of the amount of evidence, typically calculated as $c = n / (n + k)$, where $n$ is the amount of evidence and $k$ is a system parameter representing the "weight" of the prior.

By tracking confidence separately, NARS can distinguish between "I don't know because I have no data" (low confidence) and "I don't know because the data is perfectly balanced" (high confidence).

Case Study: The Evolution of Belief

Wang illustrates the NARS revision rule using a logical example involving the flight capabilities of birds:

  1. Initial Observation: A system observes several doves that fly. It forms a belief that "birds fly" with high frequency ($f$) but low confidence ($c$), as the sample size is small.
  2. Evidence Accumulation: The system then observes several swans that fly. In a Bayesian system, the probability might remain unchanged. In NARS, the frequency stays high, but the confidence increases. The system is now more certain of its uncertainty.
  3. Conflicting Evidence: The system encounters a penguin—a bird that does not fly. NARS merges this new, conflicting evidence with the old. The frequency of the belief "birds fly" drops, but the total confidence increases again. The system has a more nuanced, stable understanding because it is drawing from a larger, more diverse pool of evidence.

Conclusion

While moving away from the Kolmogorov axioms of probability requires a ground-up rebuild of logical operators, Wang argues the trade-off is necessary for general intelligence. By representing the "weight" of evidence through confidence, NARS provides a mechanism for belief revision that is more resilient and adaptive than standard Bayesian updating. Future work remains to compare NARS's computational overhead against other multi-valued systems like Dempster-Shafer theory, but the theoretical distinction between propagation and revision offers a vital path forward for AI development.

Summary updated

BibTeX
@inproceedings{wang1993belief,
  author = {Pei Wang},
  title = {{Belief Revision in Probability Theory}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}
Figure