Deep Latent Variable MFG Model

• The model augments standard MFGs by introducing a latent probabilistic context that indexes families of games with heterogeneous reward specifications while preserving intra-context homogeneity.
• It employs deep networks to parameterize context-conditioned rewards, adaptive sampler policies, and a context encoder, enabling joint learning from trajectory data through adversarial and mutual-information based objectives.
• Empirical results demonstrate improved policy matching, lower variance in expected returns, and practical benefits in applications such as adaptive taxi pricing using real-world data.

A deep latent variable mean-field game (MFG) model augments a standard entropy-regularized discrete-time MFG with a latent probabilistic context variable $z$ that externally indexes a family of MFGs sharing the same state space, action space, transition structure, and initial population distribution while differing in reward specification. In the formulation proposed for meta-inverse reinforcement learning (meta-IRL), the objective is to recover context-conditioned rewards from demonstrations drawn from heterogeneous and unknown objectives without modifying the underlying MFG approximation itself. Deep networks parameterize the reward $f_\omega(s,a,m_t,z)$, the context encoder $q_\psi(z\mid \tau)$, and an adaptive sampler policy $\pi_\theta$, so that reward inference and context inference are learned jointly from trajectory data (Chen et al., 4 Sep 2025).

1. Formal setting and equilibrium structure

The model is posed for a continuum of homogeneous agents interacting through the population distribution, or mean field. The local state space $S$ and action space $A$ are finite, and time is discrete, $t=0,1,\dots,T$. At time $t$, an individual agent occupies state $s\in S$, selects action $a\in A$, and experiences dynamics governed by a transition kernel

$f_\omega(s,a,m_t,z)$0

where $f_\omega(s,a,m_t,z)$1 denotes the population distribution over states.

In the latent-context formulation used in the meta-IRL model, the transition function is assumed independent of $f_\omega(s,a,m_t,z)$2 for simplicity, whereas the reward depends on $f_\omega(s,a,m_t,z)$3. The reward is written as $f_\omega(s,a,m_t,z)$4 for the underlying model and as $f_\omega(s,a,m_t,z)$5 for the learned reward parameterization. Policies are entropy-regularized and context-conditioned: $f_\omega(s,a,m_t,z)$6

Given a mean-field flow $f_\omega(s,a,m_t,z)$7 and a policy flow $f_\omega(s,a,m_t,z)$8, the conditional trajectory distribution is

$f_\omega(s,a,m_t,z)$9

Consistency between policy and population evolution is enforced by the discrete-time McKean–Vlasov recursion

$q_\psi(z\mid \tau)$0

The equilibrium notion is an entropy-regularized mean-field Nash equilibrium. A pair $q_\psi(z\mid \tau)$1 is an equilibrium if, first, $q_\psi(z\mid \tau)$2 maximizes the entropy-regularized expected return given $q_\psi(z\mid \tau)$3, and second, $q_\psi(z\mid \tau)$4 is generated by $q_\psi(z\mid \tau)$5 through the McKean–Vlasov recursion. The corresponding control problem is

$q_\psi(z\mid \tau)$6

A frequent misconception is to equate the latent context with heterogeneity inside a single MFG. In this model, homogeneity is preserved within each context $q_\psi(z\mid \tau)$7: agents in a fixed context share the same $q_\psi(z\mid \tau)$8, $q_\psi(z\mid \tau)$9, transition structure, and reward functional form. Heterogeneity enters across contexts, not within them (Chen et al., 4 Sep 2025).

2. Latent probabilistic context and generative interpretation

The latent variable $\pi_\theta$0 is a probabilistic context variable. In the reported experiments it is discrete, for example $\pi_\theta$1, with prior $\pi_\theta$2. The dataset of expert demonstrations is modeled as a mixture over contexts: $\pi_\theta$3 where $\pi_\theta$4 denotes the expert equilibrium under context $\pi_\theta$5.

Under a maximum-entropy style formulation, the conditional trajectory model is written in energy-based form: $\pi_\theta$6 The graphical structure is straightforward: $\pi_\theta$7 influences the reward and the policy; the policy, together with dynamics and the consistency recursion, determines the mean-field flow; and the resulting equilibrium generates trajectories.

This external placement of $\pi_\theta$8 is structurally important. The method does not add agent types inside the MFG itself. Instead, it places contexts over a family of MFGs. This preserves the standard MFG template while allowing demonstrations from different tasks to be pooled. A plausible implication is that standard existence and uniqueness arguments for entropy-regularized equilibria can still be applied context-wise, because the latent variable indexes distinct reward-conditioned games rather than altering the mean-field approximation.

The model is therefore best viewed as a meta-IRL construction over structurally similar tasks. All tasks share the same $\pi_\theta$9, $S$0, $S$1, and initial mean field $S$2, but differ in rewards through $S$3, and therefore differ in equilibrium policy and equilibrium population flow (Chen et al., 4 Sep 2025).

3. Objective function, intractability, and adversarial reformulation

A natural starting point is a marginal-likelihood or ELBO-based treatment. For task datasets $S$4, a standard ELBO would take the form

$S$5

However, this is intractable in the present setting because the mean-field flow $S$6 depends on the equilibrium policy and reward parameters, and the partition function of the energy-based trajectory model is unknown.

The method therefore adopts a different tractable objective, inspired by mutual-information regularization. The optimization problem minimizes the KL divergence between expert conditional trajectory distributions and model conditional trajectory distributions while maximizing mutual information between $S$7 and trajectories: $S$8 The second term operationalizes the context objective by approximating $S$9 with $A$0, thereby making the mutual information $A$1 tractable up to an additive constant.

To address the entanglement between policy, population flow, and normalization, the method constructs a $A$2-dependent empirical estimate of the expert mean field: $A$3 This yields the conditional model

$A$4

Sampling from this model is handled adversarially through a discriminator

$A$5

If the adaptive sampler $A$6 is optimized against this discriminator with reward signal $A$7, the resulting trajectories match $A$8. This sampling-equivalence lemma is the technical device that bypasses direct normalization and direct sampling from the energy-based model (Chen et al., 4 Sep 2025).

4. Learning procedure, gradient structure, and implementation

The learning procedure alternates among reward learning, context inference, and policy optimization. The encoder $A$9 performs amortized inference from trajectories to contexts and is used both to infer contexts and to build the empirical mean field $t=0,1,\dots,T$0. For discrete $t=0,1,\dots,T$1, categorical sampling is used; no Gaussian reparameterization is required in the reported experiments.

The optimization involves alternating updates of $t=0,1,\dots,T$2 and $t=0,1,\dots,T$3, together with an inner loop for the adaptive sampler $t=0,1,\dots,T$4. For fixed $t=0,1,\dots,T$5, one computes $t=0,1,\dots,T$6 from expert data, trains the discriminator and sampler alternately, and then updates $t=0,1,\dots,T$7 and $t=0,1,\dots,T$8 using the derived gradient estimators for the KL term and the information term. The reported gradient expressions use

$t=0,1,\dots,T$9

and the paper gives unbiased estimators for the relevant derivatives with respect to $t$0 and $t$1.

At high level, the training loop is:

1. sample expert mini-batches;
2. infer contexts $t$2;
3. estimate $t$3;
4. generate trajectories under $t$4;
5. update $t$5 using the KL and information terms;
6. update $t$6 using the adversarial objective and the information term;
7. update $t$7 to maximize entropy-regularized returns under $t$8.

The principal network specifications reported for the method are as follows.

Component	Specification	Role
$t$9	4-layer MLP; two hidden layers of size 64; Leaky ReLU; Adam; learning rate $s\in S$0	Context-conditioned reward
$s\in S$1	Per-time-step 5-layer MLP; two hidden layers of size 64; third hidden layer size $s\in S$2; Softmax; Leaky ReLU; Adam	Adaptive sampler policy
$s\in S$3	Categorical encoder with Softmax	Context inference

States and actions are represented by one-hot encodings, and mean fields $s\in S$4 are represented as vectors in $s\in S$5. Estimating $s\in S$6 requires scanning expert trajectories once per training step, with cost $s\in S$7 per mini-batch. In the simulated experts, equilibrium computation uses fixed-point iteration with backward induction for $s\in S$8 and McKean–Vlasov updates for $s\in S$9, terminating when the mean-squared difference across successive mean-field iterates is at most $a\in A$0 over all states and times (Chen et al., 4 Sep 2025).

5. Empirical domains, baselines, and reported performance

The reported simulated domains are VIRUS, MALWARE, and INVEST. In VIRUS, agents choose distancing versus going out while infection spreads through the mean field. In MALWARE, the state tracks health or infection level and the action is intervene versus do-nothing, with reward penalizing infection and adding intervention cost coupled to the population average health. In INVEST, firms invest to improve product quality, and both dynamics and rewards depend on average market quality. For each context $a\in A$1, expert equilibria are computed, trajectories of length $a\in A$2 are sampled, and $a\in A$3 is drawn uniformly from $a\in A$4.

The baselines are PLIRL, a centralized population-level IRL method, and MF-AIRL, a mean-field AIRL method without context. The evaluation metrics are policy deviation,

$a\in A$5

and expected return difference under the ground-truth reward $a\in A$6.

The reported outcome is that PEMMFIRL achieves minor deviations from expert policies and small return gaps, whereas PLIRL and MF-AIRL show large deviations due to inability to handle heterogeneous contexts. The method also exhibits low variance across runs, with approximately $a\in A$7 lower variance in expected returns and policy deviations, while incurring acceptable training-time overhead for $a\in A$8.

The real-data application is a spatial taxi-ride pricing problem using the NYC Yellow Taxi dataset on the Queens subset. The city area is partitioned into a $a\in A$9 grid over longitude $f_\omega(s,a,m_t,z)$00 and latitude $f_\omega(s,a,m_t,z)$01 with $f_\omega(s,a,m_t,z)$02 resolution. The trajectory horizon is $f_\omega(s,a,m_t,z)$03 steps, corresponding to $f_\omega(s,a,m_t,z)$04 minutes per step. The context $f_\omega(s,a,m_t,z)$05 is used to capture driver preferences, such as trip-distance sensitivity. The reward is specified as

$f_\omega(s,a,m_t,z)$06

where $f_\omega(s,a,m_t,z)$07 is a learned profit function depending on $f_\omega(s,a,m_t,z)$08, $f_\omega(s,a,m_t,z)$09, and $f_\omega(s,a,m_t,z)$10 is the empirically set price multiplier per origin region.

For taxi pricing, the learned context-conditioned pricing improves drivers’ average profit at small passenger decay rates without introducing extra demand. The reported figures are:

• for $f_\omega(s,a,m_t,z)$11: passenger decay $f_\omega(s,a,m_t,z)$12, average profit $f_\omega(s,a,m_t,z)$13, corresponding to $f_\omega(s,a,m_t,z)$14 per ride;
• for $f_\omega(s,a,m_t,z)$15: passenger decay $f_\omega(s,a,m_t,z)$16, average profit $f_\omega(s,a,m_t,z)$17, corresponding to $f_\omega(s,a,m_t,z)$18 per ride.

Qualitative policy maps are reported to favor higher-profit areas given inferred contexts and to exhibit spatial patterns consistent with supply-demand distributions (Chen et al., 4 Sep 2025).

6. Limitations, extensions, and relation to other latent-variable MFG formulations

Several limitations are explicit. The prior $f_\omega(s,a,m_t,z)$19 is unknown, and the method relies on $f_\omega(s,a,m_t,z)$20 both to generate synthetic samples of $f_\omega(s,a,m_t,z)$21 and to construct $f_\omega(s,a,m_t,z)$22, which introduces approximation bias when $f_\omega(s,a,m_t,z)$23 is inaccurate. The reported model uses discrete $f_\omega(s,a,m_t,z)$24 and a fixed number of contexts; extending to continuous $f_\omega(s,a,m_t,z)$25 would require reparameterization and additional regularization. The empirical mean-field estimate $f_\omega(s,a,m_t,z)$26 depends on sufficient and representative demonstrations per context. In more general MFGs, equilibrium uniqueness may fail, so equilibrium selection mechanisms may be required. For high-dimensional continuous state or action spaces, the AIRL-style discriminator and sampler training may become more difficult.

The stated future directions are continuous latent contexts such as $f_\omega(s,a,m_t,z)$27, online adaptation in which $f_\omega(s,a,m_t,z)$28 infers $f_\omega(s,a,m_t,z)$29 on the fly and policies adapt with few-shot updates, theoretical work on identifiability of reward components and context recovery, richer dynamics in which $f_\omega(s,a,m_t,z)$30 also modulates the transition law $f_\omega(s,a,m_t,z)$31, and applications to traffic signal control, power-grid demand response, and social-media dynamics.

A related but distinct latent-variable perspective appears in deep-learning solvers for MFGs formulated through forward-backward stochastic differential equations with jumps. In that literature, the backward processes $f_\omega(s,a,m_t,z)$32, $f_\omega(s,a,m_t,z)$33, and $f_\omega(s,a,m_t,z)$34 are treated as latent variables encoding costate information: $f_\omega(s,a,m_t,z)$35 is the Lagrange multiplier on the state dynamics, $f_\omega(s,a,m_t,z)$36 and $f_\omega(s,a,m_t,z)$37 encode sensitivities to idiosyncratic and common noise, and $f_\omega(s,a,m_t,z)$38 encodes jump-response costate. These latent variables are learned by neural networks and induce controls through first-order optimality conditions in smart-grid demand-side management models with Cox-process activations. This suggests a broader usage of “latent variable” in MFG research: in one line of work, the latent object is an external probabilistic context over a family of MFGs; in another, it is the hidden adjoint structure of a continuous-time equilibrium computation (Alasseur et al., 2024).

Within that broader landscape, the deep latent probabilistic context model is distinguished by its specific meta-IRL purpose: it uses latent contexts to separate heterogeneous reward structures across demonstrations while preserving the original mean-field approximation and learning context-conditioned rewards, policies, and empirical mean-field surrogates jointly (Chen et al., 4 Sep 2025).