Bayesian Inverse Games

• Bayesian inverse games are probabilistic models that infer hidden strategic parameters from observed multi-agent behavior using game-theoretic equilibria and Bayesian frameworks.
• They combine forward equilibrium computations with latent variable inference to recover objectives, cost functions, and beliefs from dynamic, stochastic, or differential game settings.
• These methods are applied in fields like motion planning, multiagent control, and economic modeling, offering uncertainty quantification and improved decision-making.

Bayesian inverse games are inverse game-theoretic models in which unknown aspects of a strategic interaction—objectives, payoff parameters, types, value functions, or agents’ beliefs about one another—are treated as latent variables and inferred from observed play by combining a game-theoretic forward model with priors and an observation model. In current formulations, the forward model is typically an equilibrium computation in a noncooperative dynamic game, a mixed-strategy game, a stochastic game, or a differential game, and the inverse task is expressed either as posterior inference $p(\theta \mid y)$ or as maximum likelihood / MAP estimation under a uniform prior (Liu et al., 2024, Inga et al., 2019). The same perspective now extends to high-dimensional multi-modal observations, posterior sampling in real time, and uncertainty-aware downstream planning (Jain et al., 2 Jan 2026).

1. Scope and historical placement

The modern literature uses several near-equivalent formulations for the same underlying problem. Inverse dynamic games seek “cost function parameters … which explain observed behavior of interacting players,” multi-agent inverse reinforcement learning places the same problem in stochastic games, Bayesian opponent exploitation treats the opponent’s strategy or type as the hidden parameter updated from play, and Bayesian inverse games make the latent game parameters explicit random variables with priors and posteriors (Inga et al., 2019, Lin et al., 2014, Ganzfried et al., 2016). In all of these variants, the central operation is to infer strategic structure from behavior rather than predict behavior from known structure.

A broadening of the latent variable has become a defining feature of the area. Some works infer reward or cost parameters, some infer value functions and control penalties in differential games, and some infer each agent’s estimate of every other agent’s objective rather than only the objective itself (Bianchin et al., 8 Jan 2026, Khan et al., 5 Aug 2025). This suggests that “Bayesian inverse games” is best understood as a family of probabilistic inverse problems over multi-agent strategic models rather than a single equilibrium concept or a single algorithmic template.

The field also spans several information structures. In imperfect-information games, the hidden object may be an opponent’s mixed strategy with a Dirichlet prior updated from partial observations (Ganzfried et al., 2016). In dynamic games for motion planning, the hidden object may be a goal, desired speed, or other game parameter inferred from partial state observations (Liu et al., 2024). In stochastic games, the hidden object may be the full reward tensor over states and joint actions (Lin et al., 2014, Lin et al., 2018).

2. Probabilistic formulation

A canonical Bayesian inverse-game formulation specifies a prior over game parameters, a forward game solver, and an observation model. In dynamic games with partial observations $y$, one representative formulation is

$$b(\theta) := p(\theta \mid y) = \frac{p(y \mid \theta)\, p(\theta)}{p(y)},$$

where $p(y \mid \theta)$ is induced by the equilibrium trajectory of a generalized Nash equilibrium problem parameterized by $\theta$ (Liu et al., 2024). In this setting, the map $\theta \mapsto y$ is not direct: it is mediated by a game solver $T$ that computes an equilibrium $\mathbf{u}^*(\theta)$, after which a sensor model produces noisy observations.

A second common formulation uses maximum entropy. For inverse dynamic games with continuous-valued state and control spaces, the trajectory model is

$$p(\zeta \mid \theta_{1:N}) = \frac{\exp\left(\sum_{i=1}^{N} \theta_i^\top \mu_i(\zeta)\right)}{\displaystyle\int_{\zeta} \exp\left(\sum_{i=1}^{N} \theta_i^\top \mu_i(\zeta)\right)\, d\zeta},$$

with feature counts $\mu_i(\zeta)$ determined by the trajectory $y$0 (Inga et al., 2019). The same work notes that maximizing the likelihood under this model is equivalent to maximum a posteriori estimation under a uniform prior. This is why maximum-likelihood inverse games and explicitly Bayesian inverse games are closely related rather than disjoint categories.

In imperfect-information games, the posterior update can be written directly over opponent strategies: $y$1 where $y$2 denotes the opponent’s mixed strategy profile (Ganzfried et al., 2016). This is a particularly transparent Bayesian inverse-game instance: the hidden parameter is the opponent’s strategy, observations are public actions, and decision-making is based on the posterior mean strategy.

A further probabilistic layer appears when bounded rationality is modeled through perception noise. In inverse matrix games with Gumbel-distributed perception error, the action probabilities are softmax responses and the fixed point is a logit quantal response equilibrium. This gives a probabilistic choice model directly tied to a stochastic game-theoretic equilibrium, and the inverse problem becomes one of recovering the cost matrix that renders the desired joint strategy a unique QRE (Yu et al., 2022).

3. Equilibrium assumptions and strategic semantics

Bayesian inverse games are inseparable from the solution concept assumed in the forward model. Inverse dynamic games with continuous states and controls have been developed for three distinct cases: Pareto-efficient solutions, open-loop Nash equilibrium, and feedback Nash equilibrium (Inga et al., 2019). In the cooperative case, inference is effectively over a global sum cost; in the noncooperative cases, inference is player-specific and depends on whether strategies are open-loop or state-feedback.

Stochastic-game formulations introduce additional equilibrium families. In two-player zero-sum games, Bayesian MIRL assumes that the observed bi-policy is a minimax bi-policy and performs Bayesian inference over rewards under that assumption (Lin et al., 2014). In general-sum stochastic games, five variants have been studied: uCS-MIRL, advE-MIRL, cooE-MIRL, uCE-MIRL, and uNE-MIRL, corresponding respectively to utilitarian cooperative strategy, adversarial equilibrium, coordination equilibrium, utilitarian correlated equilibrium, and utilitarian Nash equilibrium (Lin et al., 2018). Each induces a different inverse feasible set because each imposes different rationality constraints on the observed bi-policy.

Static and noisy-action models widen the range further. Inverse matrix games with perception noise focus on games in which the players’ quantal response equilibrium is unique, and infer the cost matrix via semidefinite programs and bilevel optimization (Yu et al., 2022). In $y$3 games, inverse learning has been studied under both a Closed-form Correlated Equilibrium Maximum-Likelihood estimator and a Logit Best Response Maximum-Likelihood estimator, explicitly contrasting static equilibrium consistency with dynamic behavioral realism (Salazar et al., 15 Jan 2026).

A notable recent extension questions the standard assumption that agents share complete knowledge of one another’s objectives. Level-1 inverse games infer the true objectives $y$4 under common-knowledge assumptions, whereas level-2 inverse games infer

$y$5

meaning agent $y$6’s own objective together with its estimates of all agents’ objectives (Khan et al., 5 Aug 2025). This introduces heterogeneous internal game models into the inverse problem itself. Adjacent work on incomplete-information learning further shows that no-regret dynamics in Bayesian games converge to Bayesian coarse correlated equilibrium, which provides another equilibrium object that can constrain inverse reasoning in games with private information (Hartline et al., 2015).

4. Inference mechanisms and computational architectures

Several computational families now coexist. Maximum-entropy inverse dynamic games use feature expectation matching and maximum likelihood, with quadratic approximations of the cost around observed trajectories to handle continuous spaces (Inga et al., 2019). In this family, unbiasedness results are stated for Pareto, open-loop Nash, and feedback Nash cases under the corresponding feature expectation assumptions.

Deep latent-variable methods have become central when posterior distributions are multi-modal or observations are high-dimensional. “Auto-Encoding Bayesian Inverse Games” constructs posterior distributions of game parameters with a structured variational autoencoder whose decoder contains an embedded differentiable game solver; the model is trained from an unlabeled dataset of observed interactions and supports efficient sampling from the inferred posteriors without computing game solutions at runtime (Liu et al., 2024). A related formulation for high-dimensional multi-modal observations trains a structured variational autoencoder with an embedded differentiable Nash game solver and is designed to generate samples from the Bayesian posterior over hidden agent objectives given limited sensor observations in real time (Jain et al., 2 Jan 2026).

Mixed-strategy trajectory models provide another route to probabilistic inference. Inverse mixed strategy games with generative trajectory models represent nominal mixed strategies with a conditional variational autoencoder and then impose a KL-regularized mixed-strategy equilibrium through Bayesian Recursive Nash Equilibrium updates (Sun et al., 5 Feb 2025). This gives a distributional model over trajectories rather than a single equilibrium path, and it is explicitly motivated by uncertain agent objectives and noisy measurements.

Online Bayesian learning in differential games takes a more classical Bayesian regression form. By casting Hamilton-Jacobi-Bellman optimality conditions as linear-in-parameter residuals, the method enables sequential Gaussian updates over value-function and cost parameters, with posterior mean and covariance updated online and used for uncertainty-aware prediction from limited, noisy data without history stacks (Bianchin et al., 8 Jan 2026). In imperfect-information games, exact Bayesian updating with Dirichlet priors yields posterior opponent models and exact Bayesian best responses for a natural class of games, replacing sampling-based approximations with closed-form posterior mean computation (Ganzfried et al., 2016).

Optimization-centric inverse-game methods remain important even when they are not explicitly Bayesian. Efficient inverse multiagent learning formulates inverse game theory and inverse multiagent learning as generative-adversarial min-max optimization problems, and extends this to inverse multiagent simulacral learning, where the goal is to find parameters and an associated equilibrium that replicate the given observations in expectation (Goktas et al., 20 Feb 2025). This suggests a close interface between Bayesian inverse games and likelihood-free or risk-minimization formulations.

5. Domains of application and empirical behavior

Interactive motion planning is the most prominent application domain. In simulated driving scenarios, Bayesian inverse games have been used to infer posterior distributions over hidden goals or desired speeds and then feed these posteriors into game-theoretic planners (Liu et al., 2024). The reported effect is safer and more efficient game-theoretic motion planning, especially when compared with maximum-likelihood baselines that provide only point estimates (Liu et al., 2024). High-dimensional multi-modal observation models further report that multimodal inference reduces uncertainty when trajectory information is uninformative or unavailable, and that posterior inference enables safer downstream decision-making without sacrificing efficiency (Jain et al., 2 Jan 2026).

Differential-game formulations target adaptive interaction and shared control. Online Bayesian learning of agent behavior in differential games is demonstrated on linear-quadratic and nonlinear shared-control scenarios, with accurate prediction and quantified uncertainty highlighted as central outcomes (Bianchin et al., 8 Jan 2026). This is important because it moves Bayesian inverse games from offline identification toward sequential behavior identification and robust prediction.

Strategic reasoning under hidden information provides a second major domain. Bayesian opponent exploitation in imperfect-information games models posterior beliefs over opponent strategies and proves that, against any distribution over opponent strategies, it is optimal to respond to the mean of that distribution (Ganzfried et al., 2016). In two-player zero-sum stochastic games, Bayesian MIRL has been validated on an abstract soccer game, where the recovered rewards are evaluated by their ability to reproduce strong play and by the quality of inferred shooting-value structure (Lin et al., 2014). General-sum stochastic-game MIRL has likewise been validated on grid-world games under cooperative, coordinated, correlated, adversarial, and Nash assumptions (Lin et al., 2018).

Market and traffic applications show that inverse-game models are not confined to robotics. Inverse learning in $y$7 games examines synthetic “chicken-dare” games and traffic-interaction scenarios simulated in SUMO, comparing parameter recovery and distributional fit for CE-ML and LBR-ML models (Salazar et al., 15 Jan 2026). Efficient inverse multiagent learning reports that its approach outperforms the widely-used ARIMA method in predicting prices in Spanish electricity markets based on time-series data (Goktas et al., 20 Feb 2025). These cases suggest that Bayesian inverse games are relevant wherever equilibrium assumptions and latent incentives jointly shape observable interaction data.

6. Identifiability, model mismatch, and open problems

A persistent theme is that uncertainty quantification is not optional. Multiple works state that maximum likelihood estimation-based inverse game approaches provide only point estimates and do not quantify uncertainty, and that this can make downstream decisions overconfident or unsafe (Jain et al., 2 Jan 2026, Liu et al., 2024). Bayesian posterior representations address this by maintaining multi-modal beliefs over objectives rather than collapsing all ambiguity into a single estimate.

Identifiability depends strongly on the solution concept. In cooperative inverse dynamic games, individual player parameters can be difficult to disentangle because the recovered quantity is a global parameter associated with the sum cost; the nonlinear ball-on-beam example explicitly exhibits cooperative ambiguity (Inga et al., 2019). In general-sum stochastic games, CE and NE are not unique, which is why uCE-MIRL and uNE-MIRL require equilibrium-selection objectives rather than the hard Bayesian likelihood used for uCS, advE, and cooE (Lin et al., 2018). In level-2 inverse games, the parameterization becomes larger and the paper proves that the problem is non-convex even in linear-quadratic games (Khan et al., 5 Aug 2025).

Another recurring difficulty is model mismatch. If the equilibrium concept is misspecified, inverse estimates need not recover the strategic object of interest; this point is explicit in discussions of Bayesian reinterpretations of maximum-entropy inverse dynamic games and in comparisons between static equilibrium models and dynamic logit-response models (Inga et al., 2019, Salazar et al., 15 Jan 2026). This suggests that inverse performance is jointly controlled by payoff parameterization and by the correctness of the behavioral model.

Computational scaling is the final major constraint. Differentiable game solvers, mixed complementarity problems, and embedded equilibrium layers can be technically and numerically demanding, especially in multi-agent continuous-control settings (Liu et al., 2024, Khan et al., 5 Aug 2025). Approximate Bayesian inference, variational posteriors, and stochastic min-max optimization alleviate some of this burden, but they introduce their own biases and local-optimum issues (Goktas et al., 20 Feb 2025). Alternative equilibrium models such as quantal response equilibrium, together with uniqueness guarantees for certain matrix games, offer one route to smoother inverse objectives, though they change the semantics of rationality (Yu et al., 2022).

A broad synthesis emerges from these strands. Bayesian inverse games are no longer confined to small static games or conjugate priors. They now include maximum-entropy continuous-control models, structured latent-variable models with differentiable Nash solvers, online Bayesian identification in differential games, imperfect-information posterior updates, and higher-order inference over agents’ beliefs about others’ objectives. The common thread is the same: infer latent strategic structure from observed interaction while retaining uncertainty as part of the object being inferred.