Dynamic Cognitive Hierarchy Solution

• Dynamic Cognitive Hierarchy is a formulation that extends cognitive hierarchy theory by allowing agents to update beliefs and actions recursively over multistage games.
• It employs recursive best-response logic where agents at different sophistication levels update strategies using Bayesian inference and evolving game histories.
• DCH has practical applications in fields such as traffic dynamics, robotics, adversarial defense, and modern AI systems, demonstrating its relevance in complex decision-making.

Dynamic cognitive hierarchy solution denotes a family of formulations that preserve the recursive, level-based logic of cognitive hierarchy theory while allowing beliefs, actions, or latent parameters to evolve over stages, histories, or time. In the narrow game-theoretic sense, it refers to the dynamic cognitive hierarchy (DCH) solution for multistage games with incomplete information, where players best respond at each information set under heterogeneous and recursively truncated beliefs about others’ reasoning levels. In a broader applied sense, related dynamic cognitive hierarchy formulations appear in repeated games, receding-horizon control, traffic dynamics, adversarial defense, robotics, and recent LLM-agent systems (Lin, 2022, Li et al., 2019, Shen et al., 2024).

1. Conceptual foundations

Cognitive hierarchy theory starts from heterogeneity in strategic sophistication. In the standard construction, level-0 agents are non-strategic, and level-$h$ agents best respond under a belief distribution $g_h(k)$ over lower levels $k=0,\dots,h-1$. For symmetric two-player, two-action games, this yields an explicit recursive algorithm for computing level-dependent behavior from the payoff matrix

$$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$

with level-0 cooperating with probability $1/2$ (Gracia-Lázaro et al., 2016).

The dynamic extension generalizes this logic to multistage games. In DCH, level-0 players uniformly randomize at every information set, whereas level-$k$ players believe that all others have levels below $k$, best respond at every information set, and update beliefs about others’ levels and types as the game unfolds using Bayes’ rule. The resulting solution is unique by recursive construction. This directly relaxes the mutual consistency of beliefs required by sequential equilibrium (Lin, 2022).

An epistemic foundation is provided by the directed-rationalizability approach based on restrictions $\Delta^\kappa$ on beliefs. In this formulation, level-$k$ is interpreted as an information type rather than a direct cap on strategic sophistication. The main result is that, in generic static games, the CH solution generically coincides with $\Delta^\kappa$-rationalizability, and in generic multistage games the DCH solution generically coincides with the behavioral consequence of rationality, common strong belief in rationality, and transparency of dynamic $g_h(k)$0 (Liu, 2024).

2. Recursive construction and belief updating

In the canonical two-player, two-action setting, the recursive structure is explicit. A player cooperates whenever

$g_h(k)$1

where $g_h(k)$2 is the believed fraction of cooperators. This induces the threshold

$g_h(k)$3

For the normalization $g_h(k)$4 and $g_h(k)$5,

$g_h(k)$6

In the Snowdrift Game, level-1 assumes $g_h(k)$7 and therefore cooperates iff $g_h(k)$8. Level-2 and higher recursively construct $g_h(k)$9 from their beliefs about lower levels, and the number of behavioral sectors in the $k=0,\dots,h-1$0 parameter space increases rapidly with the cognitive level (Gracia-Lázaro et al., 2016).

The epistemic formulation expresses the same logic as restrictions on admissible beliefs. A level-$k=0,\dots,h-1$1 type only places support on lower opponent types, treats level-0 behavior as uniform over available actions, and assigns lower-level type probabilities according to a truncated distribution: $k=0,\dots,h-1$2 Under this view, finite depth is not imposed as an exogenous computational bound; it is induced by the information-type restriction itself, and the effective strategic sophistication of each type is endogenously determined (Liu, 2024).

Dynamic applications replace static lower-level heuristics with online inference. In autonomous decision-making for interactive traffic, the ego agent maintains a Bayesian belief $k=0,\dots,h-1$3 over the environment’s cognitive level, where

$k=0,\dots,h-1$4

and solves a receding-horizon optimization that maximizes expected cumulative reward subject to chance constraints. In generalized dynamic cognitive hierarchy for driving, the non-strategic layer is implemented by automata strategies, the strategic layer includes level-$k=0,\dots,h-1$5 and equilibrium-based reasoning, and the robust layer chooses

$k=0,\dots,h-1$6

thereby planning against a set of model/type combinations consistent with observed behavior (Li et al., 2019, Sarkar et al., 2021).

3. Structural properties and departures from equilibrium invariance

A central theoretical theme is that dynamic cognitive hierarchy retains recursive best-response structure but acquires new representational sensitivities. In the Snowdrift Game, any game in the region $k=0,\dots,h-1$7 with $k=0,\dots,h-1$8 and $k=0,\dots,h-1$9 can be parametrized by the class

$$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$0

Two Snowdrift games with the same $$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$1 are equivalent in the sense that any player takes the same action in both, while games of class $$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$2 and $$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$3 are action-reversed. At the same time, the solution complexity grows with cognitive level because higher-order beliefs partition the payoff space into increasingly many sectors (Gracia-Lázaro et al., 2016).

Relative to sequential equilibrium, the decisive departure is the relaxation of mutual consistency. DCH therefore violates invariance under strategic equivalence: two games with the same reduced normal form can induce different behavior because level-0 randomization depends on the extensive-form representation, and this difference propagates upward through the hierarchy. In laboratory tests using two strategically equivalent versions of the dirty-faces game, behavior differed significantly across representations, and the direction of the difference aligned with DCH. One implication is that implementing a dynamic game experiment in reduced normal form, via the strategy method, can distort behavior (Lin, 2022).

The epistemic account clarifies that this representational sensitivity is not an ad hoc anomaly but a consequence of transparent restrictions on beliefs. The same framework also connects CH and DCH to Bayesian equilibrium: the CH solution can arise as a Bayesian equilibrium in an elaborated game whose types and prior encode the hierarchy of beliefs specified by $$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$4 (Liu, 2024).

4. Sequential control and planning

In dynamic control settings, the solution is embedded in a stochastic game or predictive-control loop. One formulation represents the interaction as the six-tuple

$$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$5

with deterministic dynamics

$$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$6

reward functions $$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$7, and temporal hard constraints. The ego agent then solves a finite-horizon receding-horizon problem,

$$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$8

subject to chance constraints on future safety. The environment’s level-$$\begin{bmatrix} & C & D\ C & R & S\ D & T & P \end{bmatrix},$$9 policy is represented by a softmax over a $1/2$0-function, and Bayesian filtering updates beliefs about the environment’s cognitive level online. In simulation, this architecture produced distinct behaviors in unsignalized intersection, highway overtaking, and highway forced merging scenarios, with the ego vehicle yielding or proceeding depending on the inferred human-driver level (Li et al., 2019).

The generalized dynamic cognitive hierarchy framework for driving expands the architecture into three layers: non-strategic, strategic, and robust. Its level-0 behavior is not a static heuristic but a finite automaton representing accommodating or non-accommodating driving styles, each governed by safety aspiration thresholds. Bounded rationality is represented by satisficing concepts rather than only by noisy response: Safety-Satisficing Perfect Equilibrium (SSPE) and Maneuver-Satisficing Perfect Equilibrium (MSPE) define indifference classes relative to equilibrium actions. Level-1 agents update sets of opponent type parameters consistent with observed actions, and the robust layer plans against heterogeneous populations without assuming common knowledge. Evaluation on two large naturalistic datasets and in simulated critical traffic scenarios showed that automata strategies are well suited for level-0 behavior in a dynamic level-$1/2$1 framework and that the robust response is effective for planning in mixed populations of strategic and non-strategic reasoners (Sarkar et al., 2021).

5. Repeated interaction, population dynamics, and adversarial robustness

Several dynamic cognitive hierarchy models focus on adaptation across rounds rather than within a single extensive form. In the Snowdrift Game, an evolutionary dynamics model updates an agent’s cognitive level or belief distribution when the current payoff decreases, using only self-payoff memory and no global information about opponent actions. Simulations reported that average cooperation levels are sensitive to the assumed cognitive belief distribution, that the population cognitive-level distribution converges to similar unimodal forms across belief scenarios, that the class symmetry is roughly preserved, and that the anti-symmetry across the main diagonal is broken by the stochastic dynamics. In a different repeated-game setting, the penny-matching algorithm models the human opponent’s reasoning level as dynamically changing between rounds according to a Markov process conditioned on winning or losing; Bayesian learning then updates beliefs over both the current level and the transition parameters. The resulting AI beat 27 out of 30 volunteer human players (Gracia-Lázaro et al., 2016, Tian et al., 2019).

Day-to-day traffic dynamics provide a population-level version of the same idea. Travelers are partitioned into $1/2$2-step types with proportions $1/2$3, and a $1/2$4-step traveler’s beliefs about lower types are

$1/2$5

The general update rule is

$1/2$6

which yields CH-NTP and CH-Logit as two concrete dynamics. These models can exhibit multiple equilibria, one of which is the classical user equilibrium. Because multiple equilibria are present, global stability is generally intractable; local stability is instead analyzed through Jacobian conditions. Calibration to a 26-day online Braess-network experiment with 268 participants fit the observed oscillatory data reasonably well and inferred that about 10–20% of the population was pure myopic, with the remainder distributed across 1-step and 2-step reasoners (Shen et al., 2024).

Adversarial settings add a robustness layer. In $1/2$7-IPOMDP, recursive Bayesian inference over opponent types is augmented by anomaly detection based on action typicality and reward verification. When observed behavior becomes incompatible with every internal opponent model, the agent switches to an out-of-belief policy such as Minimax in zero-sum games or Grim Trigger in mixed-motive games. In the iterated ultimatum game, this reduced exploitation, with the sender’s reward ratio shrinking by $1/2$8. A related security application embeds level-1-versus-level-0 reasoning into deep reinforcement learning on attack graphs: the defender’s DQN integrates the attacker’s level-0 softmax policy into its transition model, and a theoretical lower bound shows that the defender’s value is no worse than that of standard DQN as attack-graph complexity increases (Alon et al., 2024, Aref et al., 22 Feb 2025).

6. Estimation, architectures, and contemporary AI reinterpretations

A distinct but related methodological strand studies dynamic hierarchical cognitive models in which the hierarchy is temporal rather than strategic. Neural superstatistics augments a mechanistic cognitive model with a high-level transition process for time-varying parameters: $1/2$9 Bayesian inference is performed by an amortized neural architecture using an LSTM summary network and a conditional generative network. Benchmarks against bayesloop and Stan showed comparable inferential quality together with much faster post-training inference, and the empirical message is that assuming static or homogeneous parameters hides important temporal information (Schumacher et al., 2022).

In cognitive robotics, cognitive hierarchy is also used as a formal meta-theory for integrating heterogeneous reasoning components. A cognitive node is defined as

$k$0

and an active hierarchy evolves through a five-stage process: prediction update, correction update, transition function update, utility update, and action update. This extension supports online learning of transition models and flexible planning operators, allowing symbolic planning, reinforcement learning, and state estimation to coexist in a single formally specified architecture. A broader review of integrated hierarchical reinforcement learning reaches a similar conclusion: compositional abstraction, predictive processing, and intrinsic motivation have each been implemented in isolation, but a unifying architecture remains an open challenge (Hengst et al., 2023, Eppe et al., 2022).

Recent LLM-agent work uses the language of cognitive hierarchy to control reasoning depth directly. CogRouter defines four cognitive levels—instinctive response, situational awareness, experience integration, and strategic planning—and trains step-level selection of cognitive depth through Cognition-aware Supervised Fine-tuning and Cognition-aware Policy Optimization. On ALFWorld and ScienceWorld, the reported result with Qwen2.5-7B is an 82.3% success rate with 62% fewer tokens than GRPO. In parallel, CHBench evaluates LLM strategic reasoning as a distribution over reasoning levels across fifteen normal-form games, reporting that reasoning levels are consistent across opponents, that the Chat Mechanism significantly degrades strategic reasoning, and that the Memory Mechanism enhances it. These developments do not reproduce the original DCH solution literally, but they extend the central idea of adaptive, level-structured cognition into contemporary AI systems and benchmarks (Yang et al., 13 Feb 2026, Liu et al., 16 Aug 2025).