Agent-Form Correlated Equilibrium

• AFCE is a refined equilibrium concept for extensive-form games that restricts deviations to single informed action switches at individual information sets.
• It occupies a middle ground in the deviation hierarchy between EFCE and CCE, ensuring tractability through efficient regret minimization and immediate obedience at each decision node.
• The framework supports robust computational techniques and complexity analysis, emphasizing its theoretical significance and practical applications in sequential decision-making.

Agent-form correlated equilibrium (AFCE) is a mediated-equilibrium concept for perfect-recall extensive-form games in which a mediator coordinates play through recommendations, while unilateral incentives are evaluated against restricted deviations at individual information sets rather than against arbitrary normal-form strategy remappings. In the deviation hierarchy developed for correlated play in sequential settings, AFCE corresponds to “one-shot, re-correlating” modifications and sits between extensive-form correlated equilibrium (EFCE) and coarse correlated equilibrium (CCE) (Morrill et al., 2020, Cheval et al., 15 Jul 2025). The terminology is not fully uniform across the literature: an earlier computational line described a causal-deviation notion as “extensive-form (or agent-form) correlated equilibrium,” whereas later work separates causal deviations from action deviations and uses AFCE for the latter (Dudik et al., 2012, Morrill et al., 2020).

1. Formal setting and core definitions

For an $N$-player extensive-form game $\Gamma$ with perfect recall, let $S_i$ denote the finite set of pure strategies for player $i$, let $S=\times_{i=1}^N S_i$, and let $\sigma\in\Delta(|S|)$ be a mediator’s recommendation distribution over full profiles $s=(s_1,\dots,s_N)$. In the action-deviation formulation, AFCE restricts unilateral deviations to informed action transformations. For each player $i$,

$$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$

Then $(\sigma,\{D_i^{\mathrm{act}}\}_i)$ is an AFCE if and only if, for every player $\Gamma$0 and every $\Gamma$1,

$\Gamma$2

The same framework also defines “coarse” and “blind” action deviations by forbidding conditioning on the recommended action inside the trigger information set; these yield AFCCE (Morrill et al., 2020).

A later local formulation writes the same agent-form intuition in terms of conditional obedience at information sets. Let $\Gamma$3 be an $\Gamma$4-player extensive-form game with perfect recall, let $\Gamma$5 be the pure strategies of player $\Gamma$6, let $\Gamma$7, and let $\Gamma$8 be a correlation plan. When $\Gamma$9, the mediator privately tells player $S_i$0 the recommended action $S_i$1 as soon as information set $S_i$2 is reached. In the agent-form model, each information set $S_i$3 is treated as an independent “agent”: when agent $S_i$4 receives recommendation $S_i$5, it may deviate exactly once by playing $S_i$6 instead of $S_i$7, but after that it must follow all further recommendations. Writing

$S_i$8

and

$S_i$9

$i$0 is an AFCE iff for every player $i$1, every information set $i$2, and every pair of actions $i$3,

$i$4

This formulation makes explicit the one-shot deviation interpretation (Cheval et al., 15 Jul 2025).

The earlier computational exposition uses different terminology. There, a moderator draws an entire profile $i$5 from a distribution $i$6 and reveals to each player only that player’s strategy component. A deviation is causal: it is specified by a trigger information set, a recommended action that triggers, and a pure continuation strategy for the subgame below the trigger. A distribution $i$7 is an $i$8-AFCE if $i$9 for every such deviation $S=\times_{i=1}^N S_i$0 (Dudik et al., 2012).

2. Deviation classes and the mediated-equilibrium hierarchy

The 2020 analysis organizes correlated equilibrium concepts in extensive-form games by the power of the allowed unilateral deviations. External deviations ignore the recommendation entirely and map every pure strategy to a fixed pure strategy; no-$S=\times_{i=1}^N S_i$1 incentive yields CCE. Action deviations trigger at one information set and replace only that action; no-$S=\times_{i=1}^N S_i$2 incentive yields AFCE. Causal deviations trigger at one information set and then fix a full continuation strategy; no-$S=\times_{i=1}^N S_i$3 incentive yields EFCE or EFCCE. Internal deviations allow arbitrary remapping of the entire pure-strategy vector; no internal-deviation incentive yields CE (Morrill et al., 2020).

Deviation class	Characterization	Equilibrium notion
Internal	Arbitrary remapping of the pure-strategy vector	CE
Causal	Trigger at $S=\times_{i=1}^N S_i$4, then fix a continuation strategy	EFCE / EFCCE
Action	Trigger at $S=\times_{i=1}^N S_i$5, switch one action, then re-correlate	AFCE / AFCCE
External	Ignore recommendation, play a fixed pure strategy	CCE

The ordering of deviation sets is

$S=\times_{i=1}^N S_i$6

Accordingly,

$S=\times_{i=1}^N S_i$7

At the same time, the paper exhibits counterexamples showing that some set-inclusions fail under the standard AI definitions; in particular, EFCE as used in that literature does not always imply AFCE (Morrill et al., 2020).

This hierarchy resolves a recurring source of confusion. AFCE is not the strongest mediated equilibrium in extensive-form games; normal-form CE remains strongest but is usually intractable to compute. AFCE is instead a tractable refinement of CCE that forbids local single-node deviations that re-correlate, while not covering the broader continuation changes allowed by causal or internal deviations (Morrill et al., 2020).

3. Sequential rationality and the hindsight perspective

The 2020 treatment places AFCE inside a broader hindsight-rationality framing for learning in general sequential decision-making settings. Rather than focusing exclusively on factored agent behavior at equilibrium, this framing studies correlated play induced by joint learning dynamics and evaluates performance in hindsight relative to modified behavior. Within that program, the paper re-examines mediated equilibrium and deviation types in extensive-form games, presents examples separating their strengths and weaknesses, proves that no tractable concept subsumes all others, and relates the deviation classes implemented by algorithms in the counterfactual regret minimization family to the equilibrium concepts in the literature (Morrill et al., 2020).

Sequential rationality for mediated play requires that, at every information set $S=\times_{i=1}^N S_i$8 of player $S=\times_{i=1}^N S_i$9, given that the other players follow their recommended strategies, player $\sigma\in\Delta(|S|)$0 cannot profitably switch from the recommended action $\sigma\in\Delta(|S|)$1 to another action $\sigma\in\Delta(|S|)$2 and then re-correlate thereafter. Under a belief system

$\sigma\in\Delta(|S|)$3

the condition is

$\sigma\in\Delta(|S|)$4

for all $\sigma\in\Delta(|S|)$5 and $\sigma\in\Delta(|S|)$6, where $\sigma\in\Delta(|S|)$7 is the counterfactual value for playing $\sigma\in\Delta(|S|)$8 at $\sigma\in\Delta(|S|)$9 and then following the recommendation under $s=(s_1,\dots,s_N)$0 (Morrill et al., 2020).

Because action deviations re-correlate immediately, they induce a simple reach-weighted “immediate” regret bound. A fully sequential AFCE imposes this inequality at every information set, not only at the root. This recasts AFCE as a recursively local obedience notion rather than merely a root-level mediated incentive condition (Morrill et al., 2020).

4. Separating examples and illustrative constructions

A canonical separating example is the extended battle-of-the-sexes game used to show that EFCE need not imply AFCE. In that game, player 1 privately chooses to Upgrade ($s=(s_1,\dots,s_N)$1) or NotUpgrade ($s=(s_1,\dots,s_N)$2), after which both players choose Event $s=(s_1,\dots,s_N)$3 or $s=(s_1,\dots,s_N)$4 simultaneously. The recommended joint-strategy distribution

$s=(s_1,\dots,s_N)$5

achieves perfect coordination with $s=(s_1,\dots,s_N)$6. This $s=(s_1,\dots,s_N)$7 is an EFCE, but there is an informed action deviation $s=(s_1,\dots,s_N)$8 for player 1 that always plays Upgrade at the root instead of the recommended $s=(s_1,\dots,s_N)$9 or $i$0, and then re-correlates by following the mediator’s recommendation at the event node. Under this deviation, player 1’s expected payoff becomes $i$1. Hence $i$2 is not an AFCE (Morrill et al., 2020).

This example is significant because it isolates exactly the kind of profitable local deviation that AFCE is designed to rule out. The deviation is neither an unconditional replacement of the entire strategy nor a fixed continuation plan chosen once and for all; it is a single-action switch followed by renewed compliance with the recommendation process. That structure is the distinctive agent-form incentive tested by AFCE (Morrill et al., 2020).

An earlier illustrative example, using the 2012 terminology, is the two-player job-market game. The student chooses study or no-study, then answers yes or no to a random question, and then the employer chooses hire or no-hire. A simple maximum-entropy AFCE in that presentation can be computed in a dozen boosting rounds: with probability about $i$3 the moderator recommends “study$i$4yes$i$5hire,” and with probability about $i$6 recommends “no-study$i$7no$i$8no-hire.” The student’s best deviation after “no-study” is to switch to “study” upon trigger, which has zero gain in expectation, and the employer’s best deviation never yields positive gain once the student is discouraged from deviating. When the weight vector $i$9 is used, the algorithm converges to the cooperative outcome “study$$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$0whatever$$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$1hire” with probability $$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$2, which also happens to be a perfect AFCE in that causal-deviation sense (Dudik et al., 2012).

5. Algorithms and approximation methods

For the action-deviation notion, AFCE can be computed by regret minimization over reach-weighted immediate counterfactual values at each information set. The paper’s Theorem 5.1 states that one can run a separate no-regret learner at each information set $$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$3 on payoff $$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$4, which guarantees no-action-deviation regret. In this view, vanilla CFR minimizes counterfactual external regret and converges to a CCE, but does not eliminate action-deviation incentives; the paper gives the extended Shapley example as a witness. To compute an AFCE, CFR must be modified so that each per-information-set learner minimizes reach-weighted immediate counterfactual regret, external or internal, thereby eliminating action deviations. The resulting “AF-CFR” algorithm provably converges to an AFCCE or AFCE at the same $$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$5 rate (Morrill et al., 2020).

This algorithmic perspective is central because it ties AFCE directly to the deviation class implemented by local no-regret learning. The paper’s summary states that AFCE “exactly matches the deviation class of ‘one-shot, re-correlating’ modifications in extensive-form games and admits efficient regret-minimization algorithms” (Morrill et al., 2020).

Under the earlier causal-deviation usage of AFCE, Dudík and Gordon develop a different computational route based on a multiplicative-weight update algorithm analogous to AdaBoost together with Markov chain Monte Carlo sampling. The optimization program maximizes

$$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$6

subject to

$$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$7

where $$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$8. The dual has exponential-family form

$$D_i^{\mathrm{act}} = \left\{ f_i^{I,a}:S_i\to S_i \;\middle|\; \exists I\in\mathcal I_i,\; a\in A(I)\text{ such that } f_i^{I,a}(s_i)(I')= \begin{cases} a,& I'=I,\ s_i(I'),& \text{otherwise} \end{cases} \right\}.$$9

A coordinate-descent procedure repeatedly chooses the deviation with highest expected regret and updates the corresponding dual coordinate. With MCMC sample sizes

$(\sigma,\{D_i^{\mathrm{act}}\}_i)$0

the algorithm halts after

$(\sigma,\{D_i^{\mathrm{act}}\}_i)$1

rounds with probability at least $(\sigma,\{D_i^{\mathrm{act}}\}_i)$2 and returns a distribution whose maximum regret is at most $(\sigma,\{D_i^{\mathrm{act}}\}_i)$3. The implementation uses skeletons, slice-Metropolis–Hastings transitions, and sampled regret estimation, and is described as requiring no representational assumptions beyond perfect recall and a simulator oracle (“succinct EFG”) (Dudik et al., 2012).

6. Threshold optimization and computational complexity

The 2025 complexity analysis studies the Threshold-AFCE problem. Given an extensive-form game $(\sigma,\{D_i^{\mathrm{act}}\}_i)$4 with perfect recall, a linear objective $(\sigma,\{D_i^{\mathrm{act}}\}_i)$5, and a rational threshold $(\sigma,\{D_i^{\mathrm{act}}\}_i)$6, the question is whether there exists an AFCE $(\sigma,\{D_i^{\mathrm{act}}\}_i)$7 such that

$(\sigma,\{D_i^{\mathrm{act}}\}_i)$8

The paper proves that this problem is NP-complete. NP-hardness already holds for two-player games without chance nodes. The reduction is from 3-SAT and constructs a game $(\sigma,\{D_i^{\mathrm{act}}\}_i)$9 with a spoiler and a verifier such that $\Gamma$00 is satisfiable if and only if there exists an AFCE of social welfare at least $\Gamma$01 (Cheval et al., 15 Jul 2025).

The NP upper bound uses a small-support characterization. Any AFCE $\Gamma$02 can be compressed to one whose support has size at most $\Gamma$03, where $\Gamma$04 is the number of relevant histories needed to express the AFCE constraints. One then guesses a support set $\Gamma$05, introduces variables $\Gamma$06 for the distribution and variables $\Gamma$07 for reach probabilities of relevant histories, and writes polynomially many linear constraints expressing the honest-continuation payoffs, the one-shot deviation payoffs, the AFCE inequalities

$\Gamma$08

and the threshold constraint. Feasibility of that rational linear system is in NP (Cheval et al., 15 Jul 2025).

The same paper situates AFCE in a broader landscape. For multiplayer stochastic extensive-form games with perfect recall, the Threshold problem is also NP-complete for EFCE, AFCCE, normal-form coarse correlated equilibrium, and extensive-form coarse correlated equilibrium. By contrast, the corresponding Threshold problem for normal-form correlated equilibrium is PSPACE-hard, while the Threshold problem for Nash equilibria in the same extensive-form setting is ER-complete. The paper identifies this as a complexity reversal relative to normal-form games: optimal correlated equilibria are computationally simpler than optimal Nash in normal form, but the opposite holds in extensive-form games (Cheval et al., 15 Jul 2025).

A further implication is that optimization and approximation separate sharply for AFCE. The exact threshold problem is NP-complete, whereas approximate AFCE with constant $\Gamma$09 can be computed in polynomial time by regret-minimization techniques (Cheval et al., 15 Jul 2025).