A New Axiomatization for Likelihood Gambles

Phan Hong GiangUAI 2006

TL;DR

This paper introduces a more general framework for making decisions under ambiguity, specifically when you have multiple plausible models for a situation but no prior probabilities assigned to them. By axiomatizing "likelihood gambles," the authors provide a principled way to compare actions and derive their value, even when the underlying probabilities are unknown, relying instead on observed data and likelihoods.

What Problem Does This Paper Address?

Imagine you're a doctor trying to decide on the best treatment for a patient. You have several possible diagnoses (models) for their condition, and you've gathered some test results (data). However, you don't have a strong prior belief about which diagnosis is most likely. This is a classic scenario of "ambiguity" – uncertainty about the underlying probabilities themselves, not just the outcome given a probability.

Traditional decision theory often relies on assigning prior probabilities to models. When these priors are unknown or hard to define, decision-making becomes challenging. This paper tackles this problem by focusing on "likelihood gambles," which represent actions in such ambiguous situations. The core challenge is to develop a mathematical framework that allows us to compare these actions and determine their "value" without needing explicit prior probabilities, instead leveraging the information provided by the data through likelihood functions.

Methods

The paper builds upon existing work on probabilistic gambles and decision theory, particularly the von Neumann-Morgenstern framework. The key innovation is the introduction and axiomatization of "likelihood gambles."

  • Likelihood Gambles: These are defined as collections of prospects, where each prospect pairs a normalized likelihood of a model with the utility (or another gamble) associated with that model. The normalization of likelihoods ensures that the highest likelihood gets a value of 1, abstracting away from the absolute probability of observing the data under a specific model.
  • Axiomatization: The authors propose a set of axioms (A1-A6) for preference relations on these likelihood gambles. These axioms are inspired by Jensen's axiomatization of probabilistic gambles and include properties like weak ordering, independence, and idempotence.
  • Utility Function: A crucial result is that these axioms lead to a representation theorem, showing that preferences over likelihood gambles can be represented by a utility function that maps gambles to 2-dimensional vectors. This vector represents the gamble's value in terms of its "best" and "worst" outcomes, weighted by likelihoods.
  • Ambiguity Attitude: The paper introduces the concept of "ambiguity attitude," characterized by an "ambiguity premium." This premium quantifies how a decision-maker's preferences deviate from pure risk neutrality when faced with ambiguity. They derive a pricing formula for likelihood gambles that incorporates this ambiguity attitude.

Datasets

Not discussed in the paper. The paper focuses on developing a theoretical framework and uses illustrative examples rather than empirical datasets.

Key Insights and Ideas

  • Data as the Primary Driver: The framework elegantly shifts the focus from subjective prior probabilities to objective data and likelihoods. This is powerful because it allows for decision-making even when prior beliefs are vague or non-existent.
  • Separation of Ambiguity and Risk: The model explicitly separates the decision-maker's "ambiguity attitude" (their taste for ambiguity) from their "risk attitude" (their preference for certainty vs. uncertainty given known probabilities). This is a significant conceptual advance.
  • Axiomatic Foundation for Ambiguity: By providing a set of axioms for likelihood gambles, the paper offers a rigorous and principled foundation for reasoning about decisions under ambiguity, analogous to how von Neumann-Morgenstern axioms do for decisions under risk.
  • Canonical Gambles as Building Blocks: The concept of "canonical gambles" (representing outcomes of 1 or 0 with varying likelihoods) proves to be a versatile tool. The paper shows that any likelihood gamble can be represented by an equivalent canonical gamble, simplifying comparisons.

Key Results

  • Representation Theorem: The paper establishes that a preference relation satisfying their axioms can be represented by a utility function mapping likelihood gambles to 2D vectors, capturing both likelihoods and outcome utilities.
  • Pricing Formula: Under the assumption of a constant ambiguity premium, a closed-form pricing formula for likelihood gambles is derived, which takes the form of a logistic function. This formula quantifies the value of an action given the data and the decision-maker's ambiguity attitude.
  • Generalization: The proposed axiomatization is presented as more general than previous approaches, particularly in how it handles compound gambles and the role of data.

Limitations and Future Work

  • Assumption of Constant Ambiguity Aversion: The derivation of the pricing formula relies on the assumption of a constant ambiguity premium. The paper acknowledges that this is a simplification and that ambiguity attitudes can be more complex.
  • Computational Complexity: While the framework is theoretically sound, applying it to very large sets of models or complex data might pose computational challenges.
  • Future Work: The authors suggest exploring more general ambiguity attitudes beyond constant aversion, investigating the implications of their framework for specific decision-making problems (e.g., in AI and economics), and further developing the connection between their axioms and existing decision models. They also highlight the potential for their approach to provide a common ground for decision-making when individuals have differing prior beliefs.

Summary updated

BibTeX
@inproceedings{giang2006new,
  author = {Phan Hong Giang},
  title = {{A New Axiomatization for Likelihood Gambles}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2006},
}
Figure