Introduction
In classical decision theory, a fundamental asymmetry exists between how we model uncertainty and how we model preferences. Probabilities are treated as distributions over a state space, allowing for additive reasoning over disjoint events. In contrast, utilities are typically treated as point functions—scalar values assigned to individual states. This mathematical disparity makes it difficult to reason about preferences with the same elegance and modularity that Bayesian networks brought to probabilistic reasoning.
Yoav Shoham's work proposes a paradigm shift: the bi-distribution. By treating utility as a distribution over "factors" rather than a function over states, we can achieve a perfectly symmetric mathematical framework for decision-making.
The Problem of Asymmetry
Traditional models define utility $u(s)$ for each state $s$. While we can sum the probabilities of states to find the probability of an event, we cannot meaningfully "sum" the utilities of states without invoking expected utility, which necessitates the re-introduction of probability. This asymmetry hinders the development of "utility networks"—tools that could simplify complex preference modeling in fields like electronic commerce or automated agency.
The Bi-distribution Framework
The core innovation is the introduction of factors—independent components of satisfaction. Instead of assigning a value to a state, utility is defined as a distribution over these factors.
A bi-distribution consists of:
- A probability distribution over states ($S_P$).
- A utility distribution over factors ($S_U$).
- A set of links ($L$) connecting states to the factors they satisfy.
In this model, expected utility becomes a symmetric calculation. One can either sum the utilities of factors weighted by the probability of the states they are linked to, or sum the probabilities of states weighted by the utility of the factors they satisfy.
Key Insights
Additivity of Utility
By shifting focus from states to factors, utility becomes additive. Just as the probability of an event is the sum of the probabilities of its constituent states, the utility of a set of factors is the sum of the utilities of its constituent factors.
Symmetry in Foundational Theorems
Shoham re-interprets the von Neumann-Morgenstern (vNM) theorem, demonstrating that its underlying mathematics are perfectly symmetric. The standard interpretation involves a single agent choosing between lotteries (probabilistic mixtures). However, the same math can describe a "benevolent dictator" choosing between individuals with different utility functions (utility mixtures).
Embedding and Complexity
The paper proves that any standard state-space utility function can be embedded into a bi-distribution. Crucially, it establishes complexity bounds: to represent $n$ distinct utility values, one needs between $\log_2 n$ and $n$ factors. For example, the 8 states shown in Table 1 can be represented by just 3 factors, as seen in Table 2.
Conclusion and Future Directions
While the bi-distribution framework provides a robust mathematical foundation, challenges remain in practical application. Identifying "natural factors" in real-world domains is more difficult than defining physical states. Furthermore, efficient methods for eliciting factor-based preferences from users are still needed.
This work lays the groundwork for utility networks. By defining conditional utility and utility independence within this symmetric framework, researchers can develop graphical models for preferences that mirror the power and flexibility of Bayesian networks.
Summary updated