Conditional Utility, Utility Independence, and Utility Networks

Yoav ShohamUAI 1997

TL;DR

This paper proposes a new interpretation of conditional utility and utility independence, making them direct analogues of their probabilistic counterparts. This allows for the development of "utility networks," graphical models that compactly represent utility distributions, analogous to how Bayesian networks represent probability distributions.

What Problem Does This Paper Address?

The field of Artificial Intelligence has seen a surge in interest in representing and reasoning about utility, as it's a fundamental component of decision theory alongside probability. While Bayesian networks have revolutionized probabilistic reasoning, a similar framework for utilities has been lacking. Existing approaches to utility independence and conditional utility in decision theory are not as computationally advantageous as their probabilistic counterparts, hindering the development of powerful AI systems for decision-making. The authors aim to bridge this gap by introducing a new, more direct analogy between probabilistic and utilitarian concepts.

Methods

The paper introduces a novel interpretation of conditional utility and utility independence by leveraging the concept of "utility distributions." These utility distributions are structured identically to probability distributions. The authors show that these can be viewed as a special case of additive multiattribute utility functions. This allows them to define "subjective" conditional utility and utility independence, which are direct mathematical analogues of conditional probability and probabilistic independence. Based on these new notions, the paper introduces "utility networks," graphical models that exploit these concepts to compactly represent utility distributions, similar to how Bayesian networks represent probability distributions.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Direct Analogy is Key: The core insight is that by reinterpreting conditional utility and utility independence to be direct mathematical analogues of their probabilistic counterparts, we can unlock similar computational advantages.
  • Subjective vs. Objective Conditioning: The paper distinguishes between "objective" conditional utility (updating based on world facts) and "subjective" conditional utility (updating based on preference structures), arguing the latter is more analogous to probabilistic conditioning.
  • Utility Distributions as the Foundation: The paper highlights utility distributions as a crucial concept, providing a structure for utilities that mirrors probability distributions, enabling the development of utility networks.
  • Teleology as the "Causality" for Utilities: For utility networks, the concept of "teleology" (mental causality, or how desires lead to actions) is proposed as the analogue to causality in Bayesian networks, providing an intuitive interpretation for the links in utility networks.

Key Results

The paper doesn't present quantitative results from empirical evaluations. Instead, its key results are theoretical:

  • A formal definition of "subjective" conditional utility and utility independence that are direct mathematical analogues of their probabilistic counterparts.
  • The introduction of "utility networks" as a graphical representation for utility distributions, analogous to Bayesian networks for probability distributions.
  • A conceptual framework for interpreting utility networks through the lens of teleology, providing an intuitive basis for their structure.

Limitations and Future Work

  • Intuitive Interpretation of Factors: The authors acknowledge that utility factors might not always correspond to intuitive concepts, and more experience is needed to determine their practical usability.
  • Computational Efficiency of Utility Networks: While the paper introduces utility networks, the discussion on their computational efficiency and practical application is left for future work.
  • Bi-Networks: The paper briefly mentions "bi-networks" that combine probability and utility networks, suggesting this as a promising avenue for future research, particularly for handling situations where probabilities and utilities need to be reasoned about simultaneously.
  • Conditions for Utility Distributions: The authors note that while they've defined utility distributions, finding simple, compact necessary and sufficient conditions for them (analogous to representation theorems in MAUT) is an open question.

Summary updated

BibTeX
@inproceedings{shoham1997conditional,
  author = {Yoav Shoham},
  title = {{Conditional Utility, Utility Independence, and Utility Networks}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1997},
}

Loading related content...

Figure