Abstract
This paper introduces a formal framework for determining the mixture proportion $\epsilon$ in probabilistic belief contraction. By modeling the update as an internal argumentation process between two agents—one representing the original belief state and the other the revised state—the authors provide a mathematically grounded method to calculate how much of the old belief should be retained. This approach bridges the gap between qualitative belief change and quantitative probability theory.
Introduction: The Challenge of Probabilistic Contraction
In the field of belief change, contraction is the process of removing a specific belief $h$ from an agent's knowledge base without necessarily adopting its negation. While straightforward in standard logic, contraction becomes a "mixture" problem when beliefs are represented as probabilities.
According to the probabilistic version of the Harper Identity, the resulting belief state ($P^-_h$) after contracting belief $h$ is a weighted combination of the original state ($P$) and a revised state ($P^*_{ eg h}$) where the agent believes the opposite of $h$. The formula is defined as: $$P^-_h = \epsilon P + (1 - \epsilon) P^*_{ eg h}$$
The fundamental challenge in this framework is determining the value of $\epsilon$. Until this research, there was no formal mechanism to justify the choice of this mixture proportion, leaving the degree of belief retention largely arbitrary.
The Argumentation Framework
The authors propose a novel approach by treating the contraction as an argumentation process between two hypothetical internal agents:
- Agent $\mathcal{X}$: Represents the original belief state $P$ and argues in favor of $h$.
- Agent $\mathcal{Y}$: Represents the revised state $P^*_{ eg h}$ and argues in favor of $\neg h$.
An argument is defined as a pair $(H, h)$, where $H$ is the support and $h$ is the conclusion. The "strength" or level of an argument is determined by the joint probability of its premises in the agent's current state. The agents compare their best arguments; a side is preferred if it has an argument with a higher probability level. If levels are tied, the side with a higher number of unique, logically distinct arguments wins.
Stability Taxonomy
To analyze how different belief states respond to contraction, the authors classify states into a taxonomy of stable, maximally stable, and unstable. This allows for a systematic mapping of contraction scenarios.
As shown in Table 1, the taxonomy identifies nine primary scenarios. This classification is crucial because it distinguishes between "interesting" scenarios where partial contraction is necessary and limiting cases where the agent either refuses to change or completely capitulates.
Calculating the Mixture Proportion $\epsilon$
The value of $\epsilon$ is derived from the outcome of the internal debate:
- Case I (Tie): If both agents have equally strong and numerous arguments, $\epsilon = 0.5$.
- Case II-a (Strength Difference): If one side has a higher argument level, $\epsilon$ is calculated as a ratio of the two levels.
- Case II-b (Numerical Difference): If levels are equal but one side has more arguments, a slowly decreasing log function determines the mixture. This ensures a "cautious" approach where a slight numerical advantage does not lead to a winner-takes-all outcome, allowing the losing side to retain a minimum proportion ($\alpha_{min}$) of influence.
Key Results and Discussion
The paper provides the first concrete, justifiable method for calculating the mixture proportion in the probabilistic Harper Identity. The authors mapped out five primary contraction scenarios (S1, S3, M3, U1, U2) and demonstrated that in many limiting cases, $\epsilon$ naturally trends to 0 or 1.
As illustrated in Table 2, scenarios such as contracting a tautology (S3, M3) or starting from an unstable state (U1, U2) result in extreme $\epsilon$ values. The authors identified that the "Stable-Contingent" scenario (S1)—where a stable agent removes a non-trivial belief—is the primary case where partial contraction (a true mixture where $0 < \epsilon < 1$) is necessary.
Conclusion
By modeling cognitive updates as a social negotiation, this research transforms a static probability problem into a dynamic debate. The resulting framework simplifies computational complexity by proving that arguments in these scenarios only rebut rather than undercut each other. Future work aims to explicitly link this argumentation structure to "epistemic entrenchment," potentially replacing arbitrary similarity measures with structured logical debate.
Summary updated