Classical decision theory relies on a beautiful but demanding premise: to make a rational choice, an agent must map their beliefs to precise numerical probabilities and their preferences to exact numerical utilities. Under the von Neumann-Morgenstern or Savage paradigms, the optimal choice is found by calculating a weighted average—the expected utility.
However, in many real-world scenarios, this numerical precision is an artificial luxury. An expert might easily state that "Scenario A is more likely than Scenario B" or that "Treatment X is preferable to Treatment Y," but forcing them to assign exact numbers like $P(A) = 0.63$ or $U(X) = 8.4$ introduces spurious precision.
To bridge this gap, a robust body of research has established a formal Qualitative Decision Theory. This framework proves that we can make mathematically rigorous, optimal decisions using only ordinal rankings and order-of-magnitude approximations.
Replacing Averages with Medians: The Sugeno Integral
In standard decision theory, the core operation is the weighted average. But if our scales are purely ordinal (e.g., ranking outcomes simply as poor, acceptable, or excellent), arithmetic operations like addition and multiplication are completely meaningless. How do we compute an "expected value" without numbers?
The breakthrough lies in shifting from standard arithmetic to lattice-based operations ($\min$ and $\max$). In the framework proposed in Possibility, uncertainty is modeled via possibility theory and preferences are handled purely ordinally, generalizing classical risk criteria.
To aggregate these ordinal values, Didier Dubois et al. demonstrated that the Sugeno integral serves as the qualitative analogue to the classical expectation. Instead of calculating an average, the Sugeno integral effectively computes a median.
The intuition behind why this works is elegant:
- The Bottleneck ($\min$): An outcome's contribution to a decision is bounded by its plausibility. If an excellent outcome is highly implausible, its value is throttled down to low.
- The Aggregation ($\max$): The overall value of an action is determined by the highest guaranteed level of satisfaction across the plausible states.
By using these operations, Didier Dubois et al. provided an axiomatic foundation showing that qualitative decision rules can mirror the rigor of Savage's quantitative approach. Furthermore, Hélène Fargier et al. demonstrated that these models can operate successfully without assuming commensurability—meaning the scale used for uncertainty does not even need to be directly calibrated to the scale used for preferences. Even when preferences cannot be neatly lined up from best to worst, Adriana Zapico proved that these qualitative utility functions can be extended to distributive lattices, deriving consistent optimistic or pessimistic decision formulas from partially ordered outcomes.
The Order-of-Magnitude Calculus
Pure ordinal rankings tell us that $A > B$, but they cannot capture the idea that $A$ is vastly better than $B$, or that an event is negligibly rare. When sequential planning requires a bit more resolution than simple rankings, researchers turn to an order-of-magnitude calculus.
The core idea, formalized by Nic Wilson, is to represent probabilities and utilities using logarithmic scales or powers of an infinitesimal parameter $\epsilon$ (similar to Spohn's disbelief functions). In this calculus:
- Negligible probabilities become higher powers of $\epsilon$ (e.g., $\epsilon^2$ is an order of magnitude less likely than $\epsilon^1$).
- When combining probabilities, addition reduces to a $\min$ operation because the most plausible explanation completely dominates the infinitesimal alternatives as $\epsilon \to 0$.
This calculus makes qualitative planning highly tractable. Judea Pearl utilized this framework to rigorously analyze conditional "ought" statements and qualitative planning under uncertainty. Remarkably, Judea Pearl later extended this to Qualitative MDPs and POMDPs, proving that these coarse, order-of-magnitude approximations are often sufficient to identify the exact same optimal policies as precise, fully quantified numerical models.
To scale this to complex, multi-variable systems, Radu Marinescu et al. introduced Order-of-Magnitude Influence Diagrams (OOM-IDs). By pairing order-of-magnitude calculus with structured variable elimination algorithms (like ELIM-OOM-ID), they provided a way to compute robust sequential policies using only "ballpark" qualitative data.
Proving Rationality: Representation Theorems
For qualitative decision-making to be viable in high-stakes domains like healthcare, it cannot rely on ad-hoc heuristics; it must satisfy formal rationality properties, a necessity emphasized by Francesca Toni et al.. Just as classical utility theory relies on Savage's axioms to prove that maximizing expected utility is the only rational choice, qualitative decision theory requires its own representation theorems.
A foundational milestone was achieved by Ronen Brafman et al., who provided the formal mathematical "proof of rationality" for the maximin decision criterion. They proved that if an agent's behavior satisfies a few simple, intuitive axioms—most notably closure under union—their choices can be perfectly represented as a maximin strategy.
Similarly, Phan Hong Giang et al. built a qualitative counterpart to the von Neumann-Morgenstern utility theory tailored for Spohn's epistemic beliefs, defining a "qualitative expected utility" driven by minimization and addition.
However, purely ordinal frameworks face a fundamental challenge: balancing caution and decisiveness. Didier Dubois et al. demonstrated that while the foundational Sure-Thing Principle (if an action is preferred whether an event happens or not, it should be preferred overall) can be preserved in ordinal settings, standard mechanisms for combining uncertainty and preferences can inadvertently lead to overly risky or paralyzed, indecisive behavior. To resolve this fragmentation between purely pessimistic or optimistic models, Phan Hong Giang et al. introduced a unified axiomatic framework that replaces disparate risk-attitude axioms with a single axiom of qualitative monotonicity, reconciling the two viewpoints under a single mathematical umbrella.
Generalization and Learning
The ultimate expression of this paradigm is the decoupling of decision theory from the real number line entirely. Francis C. Chu et al. introduced Generalized Expected Utility (GEU), a framework that replaces standard arithmetic with abstract algebraic operators. GEU is powerful enough to represent any preference relation using generalized domains for plausibility and utility, proving that the machinery of expected utility does not inherently require numerical payoffs.
This qualitative abstraction yields massive practical benefits in machine learning. In reinforcement learning, specifying an exact numerical reward function is notoriously difficult and error-prone. Arkady Epshteyn et al. introduced Qualitative Reinforcement Learning, a framework where an expert simply provides qualitative statements (e.g., "State A is better than State B"). The algorithm automatically bounds a set of quantitative MDPs consistent with these assertions. This drastically restricts the search space, eliminating the need for exhaustive, potentially dangerous exploration.
Ultimately, these mathematical frameworks mirror how humans naturally evaluate choices. As demonstrated by Nachshon Korem et al., human attitudes toward risk and ambiguity remain remarkably consistent whether they are dealing with quantitative monetary outcomes or purely qualitative medical outcomes. By formalizing these processes through possibilistic logic and ordinal calculi, as explored in Logical Representation and Computation of Optimal Decisions in a Qualitative Setting, qualitative decision theory provides a rigorous, mathematically sound bridge between human reasoning and automated planning.
Key Foundational Work
Possibility Theory as a Basis for Qualitative Decision Theory (1995) — contributed 5 concept elements to this direction.
From Conditional Oughts to Qualitative Decision Theory (1993) — contributed 2 concept elements to this direction.
An Order of Magnitude Calculus (1995) — contributed 2 concept elements to this direction.