LBR-ML: Logit Best Response MLE in 2x2 Games

• LBR-ML is a parametric inverse learning method that models adaptive, bounded-rational game play in repeated 2x2 settings using logit best responses.
• It employs a Markov chain framework to derive a unique stationary distribution for inferring player utility parameters and rationality levels.
• Empirical studies, including traffic and synthetic experiments, demonstrate its effectiveness in recovering decision metrics under various sample sizes.

The Logit Best Response Maximum-Likelihood Estimator (LBR-ML) is a parametric inverse learning approach for modeling strategic adaptation in repeated $2\times2$ games. Designed to infer player utility parameters and rationality levels from joint-action data, LBR-ML uniquely emphasizes stochastic, path-dependent dynamics arising from bounded-rationality, as opposed to consistency with static equilibrium concepts. This methodology formalizes the connection between behavioral game theory and statistical estimation by directly fitting the long-run stationary distribution induced by repeated logit best response updates to observed data through maximum-likelihood optimization (Salazar et al., 15 Jan 2026).

1. Game Model and Stochastic Logit Best Response

LBR-ML operates in two-player, $2\times2$ normal-form games. Each player $i\in\{1,2\}$ selects between two actions $A_i=\{a_i^1,a_i^2\}$, yielding four possible joint action profiles $A=A_1\times A_2=\{a(1),a(2),a(3),a(4)\}$ with standard lex order:

• $a(1) = (a_1^1,a_2^1)$
• $a(2) = (a_1^1,a_2^2)$
• $a(3) = (a_1^2,a_2^1)$
• $a(4) = (a_1^2,a_2^2)$

Each player’s payoff is linear in known features: $u_i^{w_i}(a) = \phi_i(a)^\top w_i$ for $w_i\in\mathbb{R}^d$. The core behavioral assumption is that agents select actions via the (Blume) logit best response: at each stage $k$, given the opponent’s last action, player $i$ chooses $a_i$ with probability

$$\sigma_i(\lambda_i, w_i)[a_i\,|\,a_{-i}(k)] = \frac{\exp\left(\lambda_i\cdot u_i^{w_i}(a_i, a_{-i}(k)) \right)}{\sum_{a_i'} \exp\left(\lambda_i\cdot u_i^{w_i}(a_i', a_{-i}(k))\right)}$$

where $\lambda_i \geq 0$ is the rationality (inverse-temperature) parameter. As $\lambda_i \to \infty$, choices concentrate on best responses; as $\lambda_i \to 0$, behavior becomes random.

2. Markov Transition Dynamics and Stationary Distribution

One-step transition probabilities form a $4\times4$ Markov chain, with

$$P(\lambda, w)_{k\ell} = \prod_{i=1}^2 \sigma_i(\lambda_i, w_i)[a_i(\ell)\,|\,a_{-i}(k)]$$

Defining

• $s_1 = \sigma_1(\lambda_1, w_1)[a_1^1\,|\,a_2^1]$, $s_2 = \sigma_1(\lambda_1, w_1)[a_1^1\,|\,a_2^2]$,
• $t_1 = \sigma_2(\lambda_2, w_2)[a_2^1\,|\,a_1^1]$, $t_2 = \sigma_2(\lambda_2, w_2)[a_2^1\,|\,a_1^2]$

the explicit transition matrix $P(\lambda,w)$ is given by:

Row/Col	$a(1)$	$a(2)$	$a(3)$	$a(4)$
$a(1)$	$s_1 t_1$	$s_1(1-t_1)$	$(1-s_1)t_1$	$(1-s_1)(1-t_1)$
$a(2)$	$s_2 t_1$	$s_2(1-t_1)$	$(1-s_2)t_1$	$(1-s_2)(1-t_1)$
$a(3)$	$s_1 t_2$	$s_1(1-t_2)$	$(1-s_1)t_2$	$(1-s_1)(1-t_2)$
$a(4)$	$s_2 t_2$	$s_2(1-t_2)$	$(1-s_2)t_2$	$(1-s_2)(1-t_2)$

For $0<\lambda_i<\infty$, $P$ is primitive, guaranteeing a unique stationary distribution $\sigma^*(\lambda, w)$, which solves

$$\sigma^* = \sigma^* P,\qquad \sum_{\ell=1}^4 \sigma^*[a(\ell)] = 1$$

Closed-form expressions are available for action marginals $x=\Pr(a_1^1),\ y=\Pr(a_2^1)$: $$x = \frac{s_2 + (s_1-s_2)t_2}{1 - (s_1-s_2)(t_1-t_2)}\,,\qquad y = \frac{t_2 + (t_1-t_2)s_2}{1 - (s_1-s_2)(t_1-t_2)}$$ Joint-action probabilities are then $$\{\sigma^*[a(1)] = xy,\ \sigma^*[a(2)] = x(1{-}y),\ \sigma^*[a(3)] = (1{-}x)y,\ \sigma^*[a(4)] = (1{-}x)(1{-}y)\}$$.

3. Likelihood-based Parameter Estimation

Given $T$ i.i.d. samples $D = \{a^{(t)}\}_{t=1}^T$ from the stationary regime, LBR-ML seeks $(\lambda,w)$ maximizing the log-likelihood: $$L_T(\lambda, w) = \prod_{t=1}^T \sigma^*(\lambda, w)[a^{(t)}]$$ or equivalently, minimizing the negative log-likelihood: $$\ell(\lambda,w) = -\sum_{t=1}^T \log \sigma^*(\lambda, w)[a^{(t)}]$$ The parameter set consists of $\lambda=(\lambda_1,\lambda_2)\ge 0$ and $w = (w_1,w_2)$. No regularizer is imposed in the basic LBR-ML, but $\ell_2$ or $\ell_1$ penalties are possible.

Optimization proceeds by alternating:

• Computing $s_1,s_2,t_1,t_2$ given $(\lambda,w)$
• Assembling $P(\lambda,w)$ and evaluating stationary $x,y$
• Calculating $\ell$ and its gradients via chain rule
• Taking an update step using an off-the-shelf optimizer (e.g., L-BFGS)

Per-iteration computational overhead is $O(T + d)$: $O(T)$ for the likelihood/gradient, $O(1)$ for the stationary distribution.

4. Identifiability and Optimization Properties

LBR-ML’s identifiability hinges on the variation of the feature maps $\phi_i(a)$ across all $a\in A$; otherwise, $w_i$ is not uniquely determined, especially since multiplicative scaling between $\lambda_i$ and $w_i$ produces equivalent behavior. Common practice is to fix $\lambda_i$ or anchor a component of $w$ for scale identification.

The loss surface $\ell(\lambda,w)$ is typically non-convex in joint parameters; thus, global convergence is not assured and multiple random restarts are necessary. Locally, smoothness and Lipschitz-gradient properties follow from the analytic structure of exponentials and rational functions.

5. Empirical Evaluation and Use Cases

Chicken–Dare Synthetic Experiment

For ground-truth parameters $w_1^*=[0.3,0.7]$, $w_2^*=[0.4,0.6]$ and sample sizes $T\in\{500,1000,2000\}$, LBR-ML achieves:

$T$	MAE	RMSE
500	0.155	0.194
1000	0.433	0.442
2000	0.131	0.176

Predicted joint-action distributions match the correlated equilibrium maximum-likelihood (CE-ML) baseline less closely at small $T$ but improve as $T$ increases.

SUMO Traffic Interaction

With $w$ (dimension 8) generated under CE assumptions, LBR-ML reports MAE $\sim$0.04–0.05, outperforming inverse compositional estimation (ICE) but trailing CE-ML (MAE $\sim$0.017). Fitted $(\lambda_1,\lambda_2) \approx (3.0,3.0)$ suggest near-deterministic agent policies.

Traffic Without Coordination

Under scenarios lacking any correlated equilibrium structure, LBR-ML with $\lambda$ estimated delivers $\approx72.6\%$ decision prediction accuracy, compared to $\approx62\%$ for fixed $\lambda=1.0$. Estimated $\lambda_1=1.0$, $\lambda_2=3.0$ reflect heterogenous bounded rationality.

6. Implementation Considerations and Guidelines

Practical guidance includes:

• Normalize feature maps $\phi_i$ for stability.
• Initialize $\lambda_i\approx1$, $w_i$ via ordinary logistic regression on marginal distributions.
• Employ L-BFGS with line-search and minimum 10 random restarts.
• Monitor $\lambda_i$: large values indicate near-deterministic best response, small values approach random play.

Identifiability issues persist if features lack variation or scale is unconstrained; fixing $\lambda_i$ or setting a reference weight component circumvents degeneracy.

7. Advantages, Limitations, and Recommended Usage

Advantages:

• Captures adaptive, path-dependent strategies without centralized correlating devices.
• Recovers both individual utility weights and agent rationality parameters.
• Interpretable even absent any static equilibrium in the generated data.

Limitations:

• Non-convexity in $(\lambda,w)$ necessitates random restarts.
• Scale between $\lambda_i$ and $w_i$ is non-identifiable unless anchored.
• Assumes data stationary under a single logit-response regime.

Recommended Usage:

• Preferable whenever emergent game interaction lacks static correlated equilibrium structure.
• Suitable for traffic modeling and other repeated-interaction domains where agent bounded rationality and adaptation are central.
• Not suitable when data closely follows a correlated equilibrium generated from a central device, where CE-ML provides superior fit.

LBR-ML provides a practical and theoretically grounded framework for inferring behaviorally meaningful models from aggregate multi-agent data, leveraging the analytic tractability of the $2\times2$ case to support robust optimization and comprehensive parameter recovery (Salazar et al., 15 Jan 2026).