Introduction
In strategic environments ranging from online auctions to high-frequency trading, we often observe the actions players take without knowing their underlying preferences or "values." For economists, understanding these hidden utilities is crucial for evaluating system efficiency or designing better market rules. This creates an "inverse problem": how do we infer what players want based solely on how they act?
Traditionally, econometrics has solved this by assuming agents are perfectly rational and operate in a Nash Equilibrium (NE). However, real-world human behavior frequently deviates from these theoretical ideals. This paper explores an alternative: adapting behavioral economics models—originally designed to predict how humans will act—into tools for estimating what humans value.
The Inverse Problem in Strategic Games
To test these methods, the study uses a dataset of 12 different "2x2" games (two players, each with two actions) from Selten and Chmura (2008). In each session, human subjects played 200 repeated rounds. To evaluate the accuracy of different estimation methods, the researcher "hides" one utility value (labeled 'X' in the payoff matrix) and attempts to recover it using the observed frequencies of play.
Behavioral Models as Estimation Tools
The paper adapts four established behavioral equilibrium models to work "backward" from data to parameters:
- Quantal-Response Equilibrium (QRE): Assumes players are noisy optimizers, choosing better actions more frequently but not exclusively.
- Action-Sampling Equilibrium (ASE): Players respond to a small mental sample of their opponent's past moves.
- Payoff-Sampling Equilibrium (PSE): Players sample their own past payoffs for each action and choose the historical winner.
- Impulse-Balance Equilibrium (IBE): Players adjust behavior based on the "regret" of missed opportunities in previous rounds.
These are compared against the standard Nash Equilibrium and a newer Quantal Regret (QR) method, which focuses on minimizing regret under noise without assuming a specific steady-state equilibrium.
Comparative Performance and "Sharpness"
The results demonstrate that behavioral models significantly outperform the traditional Nash approach. While the Nash Equilibrium had the highest Root Mean Squared Error (RMSE) of 3.41, the Quantal Regret (QR) method achieved the best overall accuracy (RMSE 2.32), followed closely by the Action-Sampling Equilibrium (ASE) at 2.39.
A key finding is the trade-off between "sharpness" and robustness. Behavioral models like ASE and PSE are "sharper"—they are more likely to produce highly accurate "hits" (estimates within ±3 units of the true value). ASE achieved a hit rate of 85.42%, compared to QR’s 81.13% and NE’s 68.87%. However, when behavioral models fail, they can fail significantly, leading to larger outliers.
Conclusion
This research bridges the gap between behavioral theory and econometric practice. It suggests that if the goal is to avoid large, catastrophic errors in preference estimation, the Quantal Regret method is the most robust choice. However, if the priority is the highest probability of an exact estimate, behavioral sampling models like ASE are superior. Future work may lie in "hybrid" models that combine the high hit rates of behavioral sampling with the robustness of regret-based frameworks.
Summary updated