TL;DR
This paper introduces a novel framework for inductive reasoning called "Ordinal Conditional Functions" (NCFs) that offers a non-probabilistic alternative to standard probability theory. NCFs provide a more direct way to model "plain belief" and its changes, offering a computationally simpler and potentially more intuitive approach for certain applications, especially those involving subjective expert judgments.
What Problem Does This Paper Address?
Traditional inductive reasoning is often formalized using probability theory, where beliefs are represented by probability assignments. However, probability theory struggles to capture the intuitive notion of "plain belief" – believing a proposition to be true, false, or neither – without resorting to extreme probabilities (0 or 1) which are often considered incorrigible. The paper argues that plain belief is corrigible and that probability 1 doesn't equate to plain belief. Furthermore, representing plain belief as a set of propositions or their conjunction doesn't adequately handle belief revision when new information contradicts existing beliefs. This paper aims to bridge this gap by proposing a formal, non-probabilistic theory of belief change that directly models plain belief and its dynamics.
Methods
The core of the proposed theory is the Ordinal Conditional Function (NCF). An NCF is a function that assigns natural numbers to possibilities (or "worlds") representing a "grading of disbelief." A possibility with a disbelief value of 0 is not disbelieved, while a value of 1 means it's disbelieved. The paper defines NCFs algebraically within a framework of possible worlds and propositions represented as subsets of these worlds.
Key concepts and operations introduced include:
- NCF Definition: Assigns natural numbers to possibilities, with specific rules for atoms of a proposition field. The value for a proposition is the minimum disbelief value across its constituent possibilities.
- Belief Representation: A proposition A is believed true if its negation has a positive disbelief value. The set of believed propositions is deductively closed.
- Belief Change (Conditionalization): The paper proposes two main ways NCFs change with new information:
- A,m-conditionalization: Updates belief based on a proposition A and a parameter 'm' representing the firmness of the new information.
- λ-conditionalization: Generalizes this to information concerning a whole field of propositions, parameterized by another NCF (λ). This is analogous to Jeffrey's generalized conditionalization in probability theory.
- Independence: Concepts of independence between fields of propositions are defined, mirroring probabilistic independence.
The paper also establishes a formal link between NCFs and probability measures, showing that NCFs are structurally homomorphic to probability theory, allowing for translation of probabilistic concepts like total probability and Bayes' theorem into the NCF framework.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Direct Modeling of Plain Belief: NCFs offer a more direct way to represent plain belief than probability theory, avoiding the issues with incorrigibility at probability 0 or 1.
- Corrigibility of Belief: The framework explicitly handles belief revision, even when new information contradicts existing beliefs, a crucial aspect of corrigible belief.
- Structural Homomorphism to Probability: The surprising structural similarity between NCFs and probability theory suggests that NCFs can leverage many of the same mathematical tools and insights, while potentially being computationally simpler.
- Computational Simplicity: For certain applications, especially those involving subjective judgments, eliciting and manipulating NCFs might be more straightforward than dealing with precise probabilities.
- Deterministic Causation: The framework's ability to handle conditional independence suggests potential for extending theories of probabilistic causation to deterministic causation.
Key Results
The paper presents several theorems and definitions that formalize the NCF framework:
- Theorem 1: Establishes fundamental laws for NCFs concerning negation (a proposition and its negation cannot both be disbelieved) and disjunction (disbelief in A or B is the minimum of their individual disbeliefs).
- Definition 2 & Theorem 1 (cont.): Defines the "A-part" of an NCF and provides a law for conjunction based on conditional NCFs.
- Definition 3 & 4: Formalize belief change through A,m-conditionalization and λ-conditionalization, analogous to probabilistic belief revision.
- Definition 5: Defines independence and conditional independence for fields of propositions within the NCF framework.
- Theorem 2: Establishes a formal connection between NCFs and non-standard probability measures, showing how probabilistic concepts translate to NCF operations (e.g., addition of probabilities corresponds to minimum of disbeliefs).
Limitations and Future Work
- Handling Statistical Data: The authors note that while NCFs are computationally simpler and good for subjective judgments, they are less directly tied to relative frequencies, making it unclear how to reasonably deal with statistical data within this framework.
- Comparison with Other Theories: The paper compares NCFs to various logics (conditional logic), plausibility measures (Shackle, Rescher, Cohen), Shafer's belief functions, and fuzzy logic. It argues that NCFs offer a more direct and complete theory of inductive reasoning, particularly in handling belief change and conditionalization, where other theories fall short or are incompatible.
- Future Work: The paper suggests that the NCF framework can be extended to deterministic causation and that techniques from probabilistic graphical models (influence diagrams, Markov fields, causal graphs) can be translated to NCF theory. The possibility of a "fuzzification" of NCF theory is also mentioned.
Summary updated