TL;DR
This paper explores how uncertainty is handled in logical arguments, distinguishing between uncertainty in premises and uncertainty in conclusions. It argues that while incorporating uncertainty into premises can lead to deductively valid arguments, this approach is computationally costly and doesn't always reflect human reasoning. The authors propose that a more practical approach involves allowing uncertainty in the inference procedure itself, even if it means sacrificing deductive validity.
What Problem Does This Paper Address?
In everyday reasoning and decision-making, we often deal with incomplete or uncertain information. This paper tackles the fundamental question of how to formally represent and reason with this uncertainty within the framework of logical arguments. Specifically, it addresses the distinction between:
- Uncertainty in premises: The initial information or evidence we start with might not be definitively true.
- Uncertainty in conclusions: The outcome of our reasoning process might itself be uncertain.
The authors note that some formal treatments of uncertainty focus solely on the premises, assuming the inference process itself is always certain (deductively valid). However, they question whether this is always the most appropriate or practical way to model human reasoning, which often involves making inferences that are not strictly truth-preserving but still useful. The paper aims to clarify the interplay between uncertainty in premises and the inference procedure, and to explore the implications for constructing rational arguments.
Methods
The paper proposes two formal schemas for representing arguments, inspired by Carl Hempel's work:
- Schema I (Uncertainty-Premise, Premises, P(Conclusion) = r): This schema represents an argument where at least one premise expresses uncertainty (e.g., a probability), and the conclusion is also a statement of uncertainty. The authors argue that in this schema, the inference procedure can be considered deductively valid if the uncertainty is properly incorporated into the premises. This often involves techniques like classical conditionalization in probability theory.
- Schema II (Uncertainty-Premise', Premise', Conclusion' / p): This schema represents an argument where the conclusion is categorical (unqualified), but the inference procedure itself is not necessarily deductive. The qualification 'p' next to the inference line indicates that the procedure is not truth-preserving. This schema is intended to capture inductive, probabilistic, or other forms of inference where the conclusion might be false even if the premises are true.
The authors explore the idea of making arguments deductively valid by incorporating all uncertainty into the premises. However, they argue that this approach has drawbacks: it's computationally expensive and doesn't align with how humans often reason. They suggest that allowing uncertainty in the inference procedure (Schema II) can be a more practical and simpler alternative, even if it means the argument is not deductively sound.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Distinction between "uncertainty about X" and "uncertainty of X": The paper highlights the crucial difference between a statement that expresses uncertainty about something (e.g., "The probability of rain is 0.5") and a statement that is uncertain in itself (e.g., "It will rain").
- Deductive Validity vs. Practical Reasoning: While it's possible to construct arguments that are deductively valid by encoding all uncertainty into the premises, this is often computationally prohibitive and doesn't reflect how humans naturally reason.
- The Role of Schema II: The paper champions the use of arguments where the inference procedure itself carries uncertainty (Schema II). This allows for categorical conclusions that are useful for decision-making, even if they are not guaranteed to be true.
- "Lottery Paradox" and Non-Conjunctivity: The paper touches upon the difficulty of combining justifications for individual statements into a justification for their conjunction. This is a significant challenge when trying to build complex arguments from simpler ones.
- Human Reasoning as an Approximation: The authors suggest that the way humans often reason might be a convenient shorthand or approximation of more complex, formally valid procedures.
Key Results
The paper doesn't present quantitative results in the typical sense of experimental evaluation. Instead, its contributions are conceptual and methodological:
- It provides a clear formalization of two distinct types of uncertain inference (Schemas I and II).
- It argues for the practical utility and cognitive plausibility of Schema II, where uncertainty resides in the inference process rather than solely in the premises.
- It demonstrates through examples (like the urn and ball problem) how deductive inference can be applied to uncertain premises to yield uncertain conclusions.
- It highlights the trade-offs between the robustness of deductive validity (Schema I) and the simplicity and practicality of probabilistic inference (Schema II).
Limitations and Future Work
- Computational Cost of Schema I: The paper acknowledges that incorporating all uncertainty into premises for deductive validity (Schema I) is computationally expensive.
- Non-Conjunctivity of Justifications: A significant challenge identified is that good reasons for individual statements do not automatically translate to good reasons for their conjunction. This is a key limitation for combining arguments.
- "Lottery Paradox" Implications: The paper suggests that this non-conjunctivity is related to the "lottery paradox" and points to this as an area for further investigation.
- Formalizing Schema II: While advocating for Schema II, the paper notes that its inferential procedures are not always valid in the classical sense and that dealing with the non-conjunctive character of such inferences is an ongoing area of research.
- Future Work: The authors suggest that exploring how to deal with the non-conjunctive character of inferences is a worthy area for future research, particularly in the context of nonmonotonic logic and statistical inference. They also hint at the need for a middle ground between overly liberal and overly dictatorial approaches to uncertainty.
Summary updated