Constructing utility models from observed negotiation actions

Angelo C. Restificar, Peter HaddawyIJCAI 2003

Abstract

This paper presents a novel approach to inferring an agent's private utility function through the observation of negotiation behaviors. By interpreting counter-offers as probabilistic gambles against a conflict outcome, we demonstrate that a neural network can learn to predict agent preferences with significant accuracy, bridging the gap between observable actions and hidden mathematical utilities.

Introduction

In negotiation theory, a "utility function" is a mathematical representation of an agent's preferences and their willingness to take risks. Typically, these functions are treated as private information; in competitive real-world negotiations, agents are unlikely to reveal their true valuations or risk tolerances honestly. Traditional methods for eliciting these utilities often rely on direct, hypothetical questioning, which is impractical in adversarial settings. This research addresses a fundamental challenge: how to construct a model of an agent’s utility using only the sequence of offers and counter-offers observed during a negotiation.

Methodology

The authors bridge the gap between observed actions and mathematical utility by leveraging Zeuthen’s strategy. They frame negotiation as a series of probabilistic gambles:

  1. The Gamble: When an agent receives an offer but chooses to make a counter-offer, they are rejecting a "sure thing" in favor of a gamble for a better deal.
  2. The Risk: This gamble carries the risk of "conflict" (no deal), which is considered the worst possible outcome.
  3. The Transformation: The authors prove that a counter-offer implies the agent's "certainty equivalent"—the point where they are indifferent between a sure thing and a gamble—lies within the mathematical interval between the current offer and the counter-offer.

These intervals are then used to train an Artificial Neural Network (ANN). The network is designed to predict whether an agent would prefer a specific offer over a gamble, effectively learning the shape of the agent's hidden utility curve through revealed preferences.

Experimental Setup

Due to the scarcity of real-world negotiation data with known underlying utility functions, the study utilized synthetic data for controlled evaluation. Two specific mathematical functions were used to represent "ground truth" behaviors: one risk-averse ($\gamma_1$) and one risk-prone ($\gamma_2$). Agents were simulated using standard "Boulware" (stubborn) and "Conceder" (flexible) strategies. The resulting dataset comprised 97 negotiation sequences totaling 477 transactions, creating a training set of 200 data points per transaction.

Results and Discussion

The core insight of this work is that every move in a negotiation serves as a data point regarding risk tolerance. A negotiator's stubbornness or flexibility signals a specific utility-to-risk ratio. The quality of this information is tied to the "certainty equivalent" interval; intervals that are approximately 50% of the total domain width provide the most useful signal, whereas very wide intervals introduce excessive noise.

Network accuracy as a function of average interval width.

As illustrated in Figure 1, the neural network's predictive power is inversely related to the interval width. For the narrowest intervals (average width of 0.24), the network correctly predicted the agent's preference 72% of the time. Even as the interval width increased to 0.31, the model maintained over 60% accuracy, significantly outperforming a random guess baseline of 50%. Statistical analysis, including right-tailed z-tests and t-tests ($p < 0.005$), confirmed that the model's learning was statistically significant, proving that negotiation actions contain recoverable utility information.

Conclusion and Future Work

While the current model assumes monotonic utility functions and relies on simulated data, it demonstrates a viable path for reverse-engineering preferences in competitive environments. Future research should focus on expanding this framework to complex multi-issue negotiations where trade-offs may be non-linear, and accounting for the cognitive biases inherent in human (non-rational) behavior. Additionally, scaling the model to symmetric scenarios where both parties are modeled simultaneously remains a key objective.

Summary updated

BibTeX
@inproceedings{restificar2003constructing,
  author = {Angelo C. Restificar and Peter Haddawy},
  title = {{Constructing utility models from observed negotiation actions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2003},
}
Figure