GaSPing for Utility

Mengyang Gu, Debarun Bhattacharjya, Dharmashankar SubramanianAAAI 2020 · OpenReview

Abstract

High-consequence decisions often require a detailed investigation of a decision maker's preferences, as represented by a utility function. Inferring a decision maker's utility function through assessments typically involves an elicitation phase where the decision maker responds to a series of elicitation queries, followed by an estimation phase where the state-of-the-art for direct elicitation approaches in practice is to either fit responses to a parametric form or perform linear interpolation. We introduce a Bayesian nonparametric method involving Gaussian stochastic processes for estimating a utility function from direct elicitation responses. Advantages include the flexibility to fit a large class of functions, favorable theoretical properties, and a fully probabilistic view of the decision maker's preference properties including risk attitude. Through extensive simulation experiments as well as two real datasets from management science, we demonstrate that the proposed approach results in better function fitting.

Introduction

In decision theory, accurately modeling human preferences is a cornerstone of effective decision support. These preferences are typically encoded in utility functions. However, the process of "eliciting" these functions from human subjects is fraught with challenges: individuals are often inconsistent, prone to fatigue, and influenced by cognitive biases.

The standard pipeline for utility modeling involves two phases: elicitation (asking questions about preferences) and estimation (fitting a mathematical model to the responses). Traditional estimation methods often rely on rigid parametric forms, which may lack the flexibility to capture complex human desires, or linear interpolation (LI), which results in "jagged" functions. These jagged functions are problematic because they lack the smoothness required to calculate derivatives—essential for determining risk attitudes via measures like the Arrow-Pratt index. Furthermore, neither approach effectively accounts for the inherent noise in human responses.

The GaSP Framework

To address these limitations, we propose a Bayesian nonparametric approach using Gaussian Stochastic Processes (GaSP). This method offers a flexible, probabilistic framework for utility estimation.

Technical Innovations

A key innovation is the use of the Matérn kernel as the covariance function. Unlike simpler kernels, the Matérn kernel ensures that the resulting utility function is twice differentiable. This mathematical property is critical for decision analysis, as it allows for the direct calculation of risk aversion measures.

Figure 1: GaSP provides a smooth, probabilistic estimation of the utility function, including uncertainty bands (shaded region) that reflect the model's confidence.

Furthermore, we demonstrate that linear interpolation is a limiting case of a GaSP (specifically, a Wiener process). By adopting a more sophisticated kernel, our approach effectively generalizes and improves upon the current state-of-the-art. To handle the small datasets typical of human elicitation, we employ an objective Bayesian framework, ensuring robust estimation even with minimal data points.

Experimental Evaluation

The GaSP model was evaluated against three distinct data sources:

  1. Synthetic Data: Simulated power utility functions with added Gaussian noise to test robustness against human inconsistency.
  2. Single-Attribute Real Data: Preferences from 48 participants regarding financial gains and losses, capturing behaviors like loss aversion.
  3. Multi-Attribute Real Data: A study involving 22 students rating university courses based on interest, teaching quality, and expected grades.

Key Results and Insights

The GaSP approach demonstrated significant improvements over traditional methods across all datasets.

  • Accuracy: In synthetic trials, GaSP's Mean Squared Error (MSE) was frequently 10 to 100 times lower than that of linear interpolation.
  • Risk Profiling: GaSP excels at identifying global curvature. In noisy environments where linear interpolation often misclassified users' risk attitudes, GaSP correctly identified whether a user was risk-averse or risk-seeking.
  • Real-World Utility: On multi-attribute course selection data, GaSP achieved an $R^2$ of 0.9813, outperforming the previous best nonlinear least squares approach (0.9391).
Figure 5: Comparison of Mean Squared Error (MSE) between GaSP and Linear Interpolation (LI). GaSP consistently achieves significantly lower error rates in noisy scenarios.

Beyond raw accuracy, the probabilistic nature of GaSP provides a full distribution of preferences. This allows decision-support systems to quantify uncertainty, which is vital for high-stakes scenarios where "knowing what you don't know" is as important as the estimate itself.

Conclusion

By moving from rigid formulas to a flexible Bayesian nonparametric framework, we provide a more robust and theoretically sound method for utility estimation. While current work focuses on estimation, future research will address the monotonicity constraint—ensuring utility functions always reflect that "more is better"—and integrate active learning to optimize the elicitation process itself.

Summary updated

Abstract

High-consequence decisions often require a detailed investigation of a decision maker's preferences, as represented by a utility function. Inferring a decision maker's utility function through assessments typically involves an elicitation phase where the decision maker responds to a series of elicitation queries, followed by an estimation phase where the state-of-the-art for direct elicitation approaches in practice is to either fit responses to a parametric form or perform linear interpolation. We introduce a Bayesian nonparametric method involving Gaussian stochastic processes for estimating a utility function from direct elicitation responses. Advantages include the flexibility to fit a large class of functions, favorable theoretical properties, and a fully probabilistic view of the decision maker's preference properties including risk attitude. Through extensive simulation experiments as well as two real datasets from management science, we demonstrate that the proposed approach results in better function fitting.
BibTeX
@inproceedings{gu2020gasping,
  author = {Mengyang Gu and Debarun Bhattacharjya and Dharmashankar Subramanian},
  title = {{GaSPing for Utility}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=lJIVcUEiqrU},
}
Figure