Preference Inference through Strategic Cost Probing

20 concepts15 papers1 proposal

In traditional preference elicitation, an observer might ask an agent a "value query": How much do you value this specific bundle of goods? While straightforward, this approach often fails in complex or competitive environments where agents may be unable to articulate their utility functions or unwilling to reveal them honestly.

A more robust alternative has emerged across economics and machine learning: revealed preference inference. Instead of asking for values, we observe choices under constraints. By strategically varying costs, taxes, or incentives, an observer can force an agent to reveal their internal trade-offs through their actions.

The Power of Demand Queries

The fundamental tool in this research direction is the demand query (or demand oracle). Unlike a value query, a demand query asks: Given these prices, which bundle would you choose to maximize your utility?

Jackson et al. established that demand queries function as powerful optimization oracles, providing a much stronger feedback signal than simple membership queries. Because a demand query forces the agent to search their own preference space based on cost, it leaks structural information about their valuation function. Badanidiyuru et al. further demonstrated that demand oracles break theoretical barriers in combinatorial optimization, offering significantly better approximation guarantees for complex valuations than traditional methods.

Probing via Strategic Incentives

If an agent's preferences are hidden, an observer can act as a "teacher" or "principal" who strategically manipulates the environment to unmask them. This is often achieved through three primary mechanisms:

  1. Taxation and Inhibition: Adiga et al. showed that by strategically "taxing" specific actions in a Boolean game, a principal can identify the boundary where a goal becomes too expensive for an agent to pursue. This reveals the agent's logical objectives by observing which goals they abandon first as costs rise.
  2. Strategic Payments: Conversely, Zhang et al. proposed learning utility functions by offering agents payments to change their behavior. By framing this as a zero-sum game between the learner and the agent, the authors showed that even agents using unknown learning algorithms will eventually reveal their preferences through their response to these financial probes.
  3. Active Policy Teaching: Parkes et al. framed this as "policy teaching," where incentives serve a dual purpose: they guide the agent toward a desired behavior while simultaneously acting as diagnostic probes to resolve uncertainty about the agent's reward function.

Learning from Observational Traces

Preference inference does not always require active intervention; it can also be performed by analyzing the "exhaust" of rational behavior.

In negotiation, Restificar et al. demonstrated that every counter-offer is essentially a choice to reject a "sure thing" in favor of a gamble for a better deal, allowing a model to reverse-engineer utility from a sequence of offers. Similarly, Sarkar et al. used the concept of rationalisability to explain human driving behavior, identifying the specific weights drivers place on safety versus progress by finding the utility parameters that make their observed maneuvers appear optimal.

Even aggregate data can be surprisingly revealing. Chen et al. proved that the aggregate demand of a market—the total quantity of goods purchased by all buyers at specific prices—contains enough information to reconstruct the individual utility functions and budgets of every buyer in that market.

Robustness and Stochasticity

A key challenge in revealed preference is that real-world agents are rarely perfectly rational or consistent. Badanidiyuru et al. addressed this by assuming agents are "low-regret" learners. If an agent chooses a more expensive action, the private reward must be high enough to justify the cost over time, allowing an observer to bound the agent's hidden margins.

Furthermore, Birge et al. moved beyond fixed utility models to stochastically revealed preferences. The authors showed that the statistical distribution of preferences across a population can be recovered solely by observing optimal actions across diverse price and budget constraints. This suggests that even when individual behavior is noisy, the underlying distribution of preferences remains identifiable through the lens of cost-based choice.

Key Foundational Work

Constructing utility models from observed negotiation actions (2003) — contributed 1 concept elements to this direction.
Preference Elicitation and Query Learning (2004) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_revealed_preference_inference_2551020e,
  title = {revealed preference inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/revealed-preference-inference/rdr_2551020e2ffd48dd9493d7e2e2b4ed98},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_revealed_preference_inference_2551020e_v1,
  title = {revealed preference inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/revealed-preference-inference/rdr_2551020e2ffd48dd9493d7e2e2b4ed98/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake