---
title: 'Stackelberg Markov Games: Models & Dynamics'
url: https://www.emergentmind.com/topics/stackelberg-markov-games
type: topic
---

# Stackelberg Markov Games: Models & Dynamics

Stackelberg Markov games are stochastic dynamic leader–follower games in which a state variable evolves in a Markovian manner and strategic asymmetry is built into the move order: a leader commits to a policy, mixed action, or dynamic rule, followers respond after observing that commitment or the realized leader action, and rewards and transitions depend on the resulting joint behavior. In the literature, the term covers several related formulations rather than a single canonical model, including finite-horizon discrete-time dynamic Stackelberg games with per-state mixed-strategy commitment [2202.04786], discounted two-player general-sum Markov games with sequential action selection [2604.04394], Bayesian typed formulations with uncertainty over attacker type [2007.10457], asymmetric-information models that become Markovian in a common-information belief state [2005.01997], and mean-field extensions in which a leader interacts with a population summarized by a distributional state [2201.05959].

## 1. Formal models and state representations

A common finite-horizon formulation is the discrete-time dynamic Stackelberg game
\[
(S,A,B,r,u,P),
\]
where \(S\) is the state space, \(A\) the leader’s action set, \(B\) the follower’s action set, \(r\) the leader reward, \(u\) the follower utility, and \(P\) the transition kernel. In the episodic layered model, the state space is partitioned as \(S=S_1\cup\cdots\cup S_H\) with \(S_1=\{s_1\}\), and transitions move only to the next layer. The leader uses a state-feedback policy \(\pi:S\to\Delta(A)\), and the follower best-responds at each visited state. This formulation is Markovian because the current state, the leader’s mixed action, the follower response, and the transition kernel determine the next-state distribution and current payoff [2202.04786].

A different but equally standard formulation is the discounted infinite-horizon tabular two-player general-sum game with finite state space \(\mathcal S\), leader action space \(\mathcal A\), follower action space \(\mathcal B\), transition kernel \(P(\cdot\mid s,a,b)\), rewards \(r^1,r^2\), and discount factor \(\gamma\in[0,1)\). The leader policy is \(\pi^1(a\mid s)\), while the follower policy conditions on both state and leader action, \(\pi^2(b\mid s,a)\). In this model the Stackelberg equilibrium is a bilevel object:
\[
\pi^{2*}(\pi^1)\in\arg\max_{\pi^2}V^2_{\pi^1,\pi^2}(s),\qquad
\pi^{1*}\in\arg\max_{\pi^1}V^1_{\pi^1,\pi^{2*}(\pi^1)}(s),
\]
and the associated fixed point can be represented through Stackelberg Q-functions [2604.04394].

Typed and asymmetric-information variants enlarge the state. A Bayesian Stackelberg Markov game is written as
\[
(P,S,\Theta,A,\tau,U,\gamma^\mathcal D,\gamma^\mathcal A),
\]
where \(\Theta\) is a collection of state-dependent distributions over follower types, \(A\) contains the defender and type-contingent attacker action sets, \(\tau\) depends on attacker type, and \(U\) contains type-specific utilities. The game is discounted and sequential, with a defender policy committed in the presence of uncertainty over which follower type is realized [2007.10457]. In partial-observation models, the strategically relevant Markov state is often a common belief \(\pi_t\) over a private physical state rather than the physical state itself, so the “Markov” property is belief-state rather than fully observed state-state Markovianity [2005.11853].

Mean-field models replace finitely many negligible followers by a population distribution. In discrete-time Stackelberg mean field games, the common state is typically a pair \((\pi_t,z_t)\), where \(\pi_t\) is a common belief on the leader’s private state and \(z_t\) is the mean-field distribution of follower states. The resulting equilibrium is a Stackelberg mean field equilibrium rather than a finite-player Markov perfect equilibrium [2201.05959].

## 2. Commitment, observability, and move order

The defining structural feature is sequential commitment rather than simultaneous play. In the finite-horizon discrete-time model, at a visited state \(s\) the leader chooses a mixed strategy \(\mu_s\in\Delta(A(s))\), the follower observes that mixed strategy and chooses
\[
b\in\arg\max_{b'\in B}\mathbb E_{a\sim\mu_s}[u(a,b')],
\]
the leader observes \(b\), and only then is a pure leader action \(a\sim\mu_s\) sampled. The leader receives \(r(s,a,b)\) and the system transitions according to \(P(\cdot\mid s,a,b)\). This ordering is explicitly different from simultaneous-move stochastic games and from formulations in which the leader’s pure action is chosen before the follower responds [2202.04786].

Other formulations use a realized-action version of the same asymmetry. In discounted tabular Stackelberg Q-value iteration, the leader chooses \(a\in\mathcal A\), the follower observes \(a\), chooses \(b\in\mathcal B\), and the state transitions according to \(P(\cdot\mid s,a,b)\). The follower’s policy therefore conditions on \((s,a)\), not just on \(s\) [2604.04394].

Asymmetric-information models change what is observed but preserve the leader–follower asymmetry. In one security formulation, the follower observes the physical state while the leader does not. The common information is the public action history
\[
h_t^c=a_{1:t-1},
\]
while the follower’s private history is
\[
h_t^f=(a_{1:t-1},x_{1:t}).
\]
The common-information state is the belief
\[
\pi_t(x_t)=P^\theta(x_t\mid a_{1:t-1}),
\]
which evolves by a Bayesian filter
\[
\pi_{t+1}=F(\pi_t,\gamma_t^f,a_t).
\]
The leader acts from common history or belief, while the follower acts from common information plus private state [2005.11853].

A common misconception is that Stackelberg Markov games differ from simultaneous-move Markov games only by equilibrium selection. The formulations above indicate a stronger distinction: the follower may observe the leader’s policy, mixed action, or realized action before moving; the follower’s feasible action set may depend on the leader’s move; and the relevant state can be a belief state rather than a publicly observed physical state. These differences change both the admissible policy classes and the form of the Bellman or fixed-point equations [2202.04786][2005.11853][2401.12437].

## 3. Equilibrium concepts

The baseline solution concept is Stackelberg equilibrium: the follower best-responds to the leader’s commitment, and the leader optimizes anticipating that response. In finite-horizon discrete-time models with a single follower, the follower best-response correspondence is often written as
\[
\varphi(\mu)=\arg\max_{b\in B}\mathbb E_{a\sim\mu}[u(a,b)],
\]
which induces a leader reward
\[
R(s,\mu)=\mathbb E_{a\sim\mu}[r(s,a,\varphi(\mu))].
\]
The leader’s benchmark is then the best policy in hindsight with respect to this induced reward [2202.04786].

Many papers adopt the **strong Stackelberg equilibrium** convention. If the follower has multiple best responses, ties are broken in the leader’s favor. In the discrete-time episodic model, if \(\tau(\mu)\subseteq B\) is the set of follower best responses, the follower is guaranteed to pick a \(b\in\tau(\mu)\) maximizing the leader’s expected reward among all best replies. The Bayesian Stackelberg Markov game model likewise uses Strong Stackelberg Equilibrium, under which each follower type chooses a deterministic best-response policy and ties among type-wise best responses are resolved optimistically for the leader [2202.04786][2007.10457].

Once multiple followers are present, the lower level is no longer a single-agent best response. In general-sum episodic Markov games with one leader and \(N\) followers, the relevant object is a **Stackelberg-Nash equilibrium (SNE)**: for a fixed leader policy \(\pi\), the followers play a Nash equilibrium \(\nu\in \mathrm{BR}(\pi)\) of the stagewise follower game induced by \(\pi\), and the leader optimizes against an equilibrium selection \(\nu^*(\pi)\) that is optimistic in the leader’s favor. The bilevel program is
\[
\max_{\pi,\nu}V_{l,1}^{\pi,\nu}(x)\qquad\text{s.t.}\qquad \nu\in \mathrm{BR}(\pi).
\]
In that paper, follower myopia is essential: followers maximize instantaneous rewards, not continuation values [2112.13521].

Under partial observability, equilibrium concepts are belief-based. The security-game RL formulation uses **Markov Perfect Stackelberg Equilibrium (MPSE)**: the leader commits to a policy over common-information beliefs, and the follower best-responds using the current private state and the same belief. In smart-grid demand response with private user storage, the paper introduces **private Markovian strategies (PMS)** and **private Markovian equilibrium (PME)**, where admissible deviations are restricted to policies measurable with respect to each player’s private local state rather than the full system state. In mean-field settings the analogous concepts are **Stackelberg mean field equilibrium (SMFE)** and, with multiple leaders, **SMFE-ML** [2005.11853][2509.12225][2209.03186].

This proliferation of equilibrium notions reflects a substantive modeling point rather than mere terminological variation. A plausible implication is that “Stackelberg Markov equilibrium” denotes a family of bilevel fixed points indexed by the policy class, observation structure, and follower coordination model.

## 4. Dynamic programming, recursive structure, and computation

When the follower response model is known, Stackelberg Markov games often admit dynamic-programming structure on the leader side. In the finite-horizon discrete-time game with linearly parameterized follower utility, if the true follower parameter \(\theta^\star\) is known, the optimal mixed strategy \(\mu_s\) at state \(s\) can be obtained by a Bellman-style bilevel program and computed by backward induction. The resulting known-utility problem reduces to linear programs across state layers, even though the follower best-responds stagewise [2202.04786].

In asymmetric-information settings, recursive decomposition typically proceeds through a common-information state. In the stochastic Stackelberg game with one leader and multiple followers, a common agent chooses prescriptions
\[
\gamma_t=(\gamma_t^l,\gamma_t^m)=\theta_t[\underline{\pi_t}],
\]
where \(\underline{\pi_t}\) is the vector of common marginal beliefs on private states. Beliefs evolve by a Bayes operator
\[
\underline{\pi}_{t+1}=\underline F(\underline{\pi_t},\gamma_t^{l,m},a_t^{l,m}),
\]
and the equilibrium can be computed by backward recursion over \(\underline{\pi_t}\) rather than by a global fixed point over full histories. Discrete-time Stackelberg mean field games apply the same idea to the public state \((\pi_t,z_t)\), where the forward evolution of the mean field and the backward value recursion together form a “master equation” [2005.01997][2201.05959].

Model-based value-iteration-style methods also exist. Stackelberg Q-value iteration in a discounted two-player general-sum tabular game updates
\[
Q_{k+1}^1(s,a,b)=r^1(s,a,b)+\gamma\sum_{s'}P(s'|s,a,b)\max_{a\in\mathcal A}Q_k^1(s',a,b_k(s',a)),
\]
\[
Q_{k+1}^2(s,a,b)=r^2(s,a,b)+\gamma\sum_{s'}P(s'|s,a,b)\max_{b\in\mathcal B}Q_k^2(s',a_k(s'),b),
\]
where the follower’s greedy response is computed from \(Q_k^2\) and the leader then anticipates that response in \(Q_k^1\). The analysis does not prove exact convergence in full generality; it proves a finite-time sup-norm bound of the form
\[
\|Q_k^1-Q_*^1\|_\infty \le \frac{6}{1-\gamma}\gamma^k+\frac{3\epsilon}{1-\gamma},
\]
with an analogous bound for the follower, so the guaranteed limit object is an \(\epsilon\)-neighborhood unless stronger assumptions make \(\epsilon=0\) [2604.04394].

More structured subclasses admit stronger tractability. In convex-concave zero-sum Markov Stackelberg games, the policy-space problem
\[
\min_{\pi_L}\max_{\pi_F\in \mathcal C(\pi_L)} J(\pi_L,\pi_F)
\]
can be solved by nested stochastic gradient descent-ascent. Under convex-concavity, smoothness, Slater, and bounded-variance stochastic oracle assumptions, the averaged policy pair is an \((\varepsilon+\delta,\delta)\)-recursive Stackelberg equilibrium after \(\tilde O(1/(\varepsilon^2\delta^2))\) gradient evaluations [2401.12437]. In the smart-grid private-information model, the lower-level follower game can be converted into an auxiliary Markov potential game and a pure PME can be computed in polynomial time, but that result depends on a specific aggregative and action-independent transition structure [2509.12225].

Computational difficulty remains central. In the no-regret learning model with unknown follower utility, the optimistic planning subproblem is a nonconvex quadratic program because the leader’s mixed strategy and the follower-utility parameter interact bilinearly in the margin constraints. The paper reports solving this subproblem with Gurobi. Thus theoretical no-regret guarantees and practical tractability need not coincide [2202.04786].

## 5. Learning with unknown dynamics, unknown follower utilities, or large populations

A substantial part of the literature concerns learning Stackelberg structure rather than assuming it known. One finite-horizon approach assumes that the leader knows transitions and its own reward but does not know the follower’s utility, except that it is linearly parameterized:
\[
u(a,b)=\langle f(a,b),\theta^\star\rangle.
\]
Observed best responses induce linear inequalities in \(\theta^\star\), allowing an optimistic version-space algorithm. Before each episode the leader computes an optimistic \(\epsilon\)-conservative policy by backward induction over states and candidate follower responses, executes that policy, observes follower actions, and intersects the version space with the halfspaces implied by those observations. With high probability, the regret is sublinear in \(T\), polynomial in \(m,n,H\) and a condition parameter \(d\), and independent of \(|S|\); the dependence on time is \(T^{1-1/p}\), where \(p\) is the feature dimension [2202.04786].

When dynamics are unknown and state information is asymmetric, learning is often belief-based. In stochastic Stackelberg security games, a common-information reduction yields a belief state \(\pi_t\), and an Expected Sarsa procedure with particle filters approximates the belief recursion. The resulting learned policy pair is claimed to form an \(\epsilon\)-MPSE, though the end-to-end RL convergence theory is partial and several proofs are omitted or sketched [2005.11853].

Bayesian type uncertainty leads to a different learning architecture. In Bayesian Stackelberg Markov games for adaptive moving target defense, the statewise equilibrium is a Bayesian Strong Stackelberg Equilibrium computed from current Q-matrices. **Bayesian Strong Stackelberg Q-learning (BSS-Q)** combines Q-learning updates with repeated solution of a Bayesian Stackelberg stage game for each state. The paper states that BSS-Q converges to the BSS of a BSMG and uses the approach to learn movement policies under unknown rewards and transitions with state-dependent priors over attacker types [2007.10457].

For simultaneous-move general-sum Markov games with one leader and multiple myopic followers, optimistic and pessimistic least-squares value iteration yield sample-efficient online and offline learning. In the online case the regret satisfies
\[
\tilde O\!\left(\sqrt{d^3H^3T}\right),
\]
while in the offline case the suboptimality scales as \(\tilde O(H^2\sqrt{d^3/K})\) under sufficient coverage. These guarantees rely on linear transition realizability, a statewise SNE oracle, and follower myopia, which reduces the lower level to stagewise normal-form equilibria [2112.13521].

Large-population learning motivates mean-field approximation and regularization. In a stationary discounted Stackelberg Markov game, alternating best-response updates can be stabilized by a softmax approximation to exact best responses and by entropy-style regularization. The paper proves existence and uniqueness of stationary Stackelberg equilibrium under continuity, uniqueness, and Lipschitz best-response assumptions with \(d_Ld_F<1\), and extends the same logic to a Stackelberg–mean-field equilibrium. The softmax approximation yields an \(O(\varepsilon)\) error bound relative to the true equilibrium and is presented as enabling scalable and stable learning without full knowledge of follower objectives [2509.16296].

## 6. Applications, variants, and limitations

The application range is broad. Security scheduling, anti-poaching patrols, and dynamic resource allocation motivate finite-horizon leader–follower models with state-dependent leader rewards and actions [2202.04786]. Adaptive moving target defense motivates Bayesian Stackelberg Markov games with uncertain attacker type and defender commitment [2007.10457]. Smart-grid demand response motivates private-state Stackelberg Markov games in which an aggregator sets prices and users make storage and demand decisions based on private local states [2509.12225]. Electricity tariff design motivates stationary Stackelberg Markov games and mean-field approximations in which a utility or commission sets time-varying rates for heterogeneous prosumers [2509.16296]. Reach-avoid control is modeled as a zero-sum Markov Stackelberg game because the follower’s feasible action set depends on the leader’s move through a safety constraint [2401.12437]. Shared-autonomy and intelligent disobedience motivate a sequential leader–follower model with asymmetric safety information, although that work is presented more as a conceptual bridge than as a fully developed Stackelberg Markov game theory [2603.20994].

Several nearby frameworks are related but not identical to standard Stackelberg Markov games. Continuous-time stochastic dynamic Stackelberg games with adapted open-loop controls treat the follower’s best response as an operator
\[
U^\star:\mathcal K_0\to \mathcal U_1
\]
between spaces of stochastic processes, and show that attention-based neural operators can approximate that best-response operator uniformly on compact sets. This is directly about operator approximation in open-loop control space rather than about Markov perfect equilibrium in state-feedback policy space [2411.09644]. Zero-sum stochastic linear-quadratic Stackelberg differential games with regime switching use a finite-state continuous-time Markov chain to modulate coefficients and derive explicit leader and follower feedback laws through coupled Riccati equations. That framework is Markovian, but it belongs to the regime-switching differential-games literature rather than the discrete-time stochastic-game literature [2408.17030].

The main limitations are structural. Many positive results require finite state and action spaces, known rewards or known transition kernels, exact observation of follower actions, strong Stackelberg tie-breaking, unique maximizers, myopic followers, or exact realizability of the follower model [2202.04786][2112.13521]. Belief-state and common-information approaches address asymmetric information, but they introduce enlarged state spaces and nontrivial fixed-point problems [2005.01997][2005.11853]. Exact tractability in private-information multi-follower settings currently depends on restrictive structure such as action-independent public transitions, aggregative coupling, and elimination of dominated actions [2509.12225]. Even when convergence is proved, the guarantee may be to an \(\epsilon\)-neighborhood rather than an exact equilibrium [2604.04394].

A second misconception is that followers in Stackelberg Markov games are always dynamically strategic. Several influential formulations instead assume stagewise or myopic follower response: the follower in the discrete-time dynamic Stackelberg game “aims to maximize its immediate expected utility,” and the followers in Stackelberg-Nash learning maximize instantaneous rewards only [2202.04786][2112.13521]. A forward-looking dynamic follower is therefore a modeling choice, not a defining property of the topic.

Taken together, the literature presents Stackelberg Markov games as a family of dynamic bilevel stochastic systems in which commitment, observation, and state evolution jointly determine the equilibrium object. The unifying theme is not a single formal tuple, but a recurring architecture: a Markov or belief-state environment, asymmetric move order, a lower-level best response or follower equilibrium, and a leader optimization problem defined over the induced response.

Source: https://www.emergentmind.com/topics/stackelberg-markov-games