TL;DR
This paper argues that Pollock's theory of defeasible reasoning, a prominent framework for artificial intelligence reasoning under uncertainty, fails to correctly handle the "lottery paradox." The authors demonstrate that this failure stems from an incompatibility between Pollock's approach and the skeptical notion of derivability, leading to an inability to draw warranted conclusions in certain common-sense scenarios.
What Problem Does This Paper Address?
Artificial intelligence systems often need to reason with incomplete or uncertain information. Nonmonotonic logics, which allow conclusions to be retracted as new information becomes available, are crucial for this. John Pollock's OSCAR framework is a notable example, aiming to provide a sophisticated account of rational reasoning. However, existing nonmonotonic logics, including Poole's, have struggled with the "lottery paradox," where one can infer that each individual ticket in a large lottery will lose, leading to a contradiction since one ticket must win. This paper investigates whether Pollock's theory of defeasible reasoning, which relies on probabilistic reasoning and defeaters, avoids this pitfall. The authors contend that it does not, and that Pollock's theory, like other default reasoning systems, is susceptible to similar paradoxes, particularly when dealing with skeptical inference (requiring a conclusion to hold in all possible scenarios).
Methods
The paper begins by introducing Reiter's default logic as a foundational concept for default reasoning, explaining its syntax and the idea of "extensions" (sets of derivable conclusions). It then discusses the problem of multiple extensions and introduces the "skeptical" approach, where a conclusion is only accepted if it holds in all extensions. The core of the paper then delves into Pollock's theory of defeasible reasoning, which uses probabilistic "prima facie" reasons and "defeaters" (reasons that cancel out prima facie reasons). The authors analyze Pollock's "collective defeat" principle, which states that if there are equally good, jointly inconsistent reasons for believing a set of propositions, none of them are justified. They then demonstrate how this principle, and Pollock's overall framework, falters when faced with scenarios analogous to the lottery paradox, particularly when propositions are "negatively relevant" (confirming one makes others more likely).
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Pollock's theory, despite its probabilistic foundation, shares the lottery paradox problem with other default logics. The paper argues that the core issue isn't just about multiple extensions but about the fundamental incompatibility of skeptical derivability with certain types of uncertain reasoning.
- The "lottery paradox" arises when jointly inconsistent defaults are equally strong. This forces a choice between accepting contradictions or remaining agnostic about plausible conclusions.
- Pollock's "collective defeat" principle, while intended to handle inconsistencies, can lead to unwarranted agnosticism. The paper shows how this principle can prevent justified beliefs even when dealing with negatively relevant propositions.
- The distinction between "positive" and "negative" relevance is crucial but not sufficient to resolve the paradox in Pollock's framework. While Pollock attempts to differentiate the lottery paradox from the "paradox of the preface" based on relevance, the paper argues his framework still struggles.
- The paper suggests that the problem lies in how Pollock's framework handles the "lotterization" of propositions. By constructing specific defeaters, the framework can be made to reject even highly probable conclusions.
Key Results
The paper demonstrates that Pollock's theory of defeasible reasoning, under a skeptical interpretation, cannot warrant conclusions in scenarios analogous to the lottery paradox. Specifically, it shows that for any proposition P that can be "lotterized" into a disjunction of propositions (p1 V ... V pn), a defeater can be constructed to prevent P from being warranted, even if each pi has a high probability. This is illustrated through mathematical derivations showing how probabilities can be manipulated to undermine warranted beliefs.
Limitations and Future Work
The authors acknowledge that their analysis highlights a fundamental limitation in Pollock's framework concerning skeptical default reasoning and its ability to handle the lottery paradox. They suggest that the problem might be resolved if Pollock's framework could be modified to better account for the probabilistic relationships between propositions, particularly when they are negatively relevant. However, they caution that such modifications might blur the distinction between the lottery paradox and the paradox of the preface, potentially undermining the framework's ability to distinguish between them. The paper concludes by questioning the claim that OSCAR offers epistemologically sophisticated reasoning due to these limitations.
Summary updated