Procedural Rational Players Explained

• Procedural rational players are defined by rules of reasoning—such as iterative elimination, higher-order beliefs, and counterfactual filtering—rather than solely by payoff optimization.
• They employ dynamic models like Monte-Carlo Tree Games and nested belief structures to account for opponents’ strategies and limited foresight in decision-making.
• Empirical and algorithmic studies demonstrate that structured reasoning processes improve predictive accuracy and strategic synthesis in both experimental and distributed-systems contexts.

Searching arXiv for recent and foundational papers relevant to procedural rational players. Procedural rational players are players whose behavior is characterized by the procedure that generates an action, rather than by outcome-based optimality alone. In this usage, rationality is specified through iterative elimination, higher-order reasoning, belief refinement, admissibility, limited-foresight search, or counterfactual elimination of impossible outcomes. Across these formulations, the common object is not merely a best response, but a player whose action survives a structured process of reasoning about opponents, observations, and feasible continuations (Chen et al., 2023, Vito, 27 Nov 2025, Turrini, 2015, Fourny, 2017).

1. Higher-order and iterative conceptions of rationality

A central formulation defines higher-order rationality through rationalizability and iterated elimination of strictly dominated strategies. Let $R_i^k(\gamma)$ be the set of strategies for player $i$ in game $\gamma$ that survive $k$ rounds of iterated elimination of strictly dominated strategies. A player $i$ exhibits $k$th-order rationality in $\gamma$ iff the player always plays a strategy in $R_i^k(\gamma)$. The nesting property

$$R_i^{k+1}(\gamma) \subset R_i^k(\gamma)\quad \forall k \in \mathbb{N}_0$$

implies that $k$-rationality entails $i$0-rationality for all $i$1. In this hierarchy, first-order rationality means not playing a strictly dominated action, second-order rationality means best responding to a belief that others are first-order rational, and the recursion continues accordingly (Chen et al., 2023).

This formulation also separates realized strategic behavior from latent strategic capacity. The distinction is expressed through $i$2, player $i$3's capacity for game $i$4, and $i$5, player $i$6's realized level in game $i$7 under environment $i$8, with

$i$9

Observed behavior therefore provides only a lower bound on true reasoning depth when beliefs about others and social preferences are not controlled (Chen et al., 2023).

A related static formulation under incomplete information uses point rationalizability as an iterative elimination procedure. Starting from $\gamma$0, the sets $\gamma$1 retain choices that remain optimal for some admissible surviving belief. Under weakly increasing differences, the surviving set is an interval,

$\gamma$2

and the limit set is

$\gamma$3

The paper proves that greater choice-parameter beliefs leads to greater optimal choices, and that the greatest and least point rationalizable choice of a player is increasing in their parameter (Sloun, 26 Jan 2025).

Taken together, these models define procedural rationality as survival under a rule-governed screening process. A plausible implication is that “rational player” denotes not only payoff maximization, but continued admissibility under recursive elimination and belief-based filtering.

2. Belief refinement, caution, and epistemic priority

In sequential games, prudent rationalizability defines a player procedurally through an iterative reduction of beliefs. Starting from $\gamma$4, the $\gamma$5th round retains strategies

$\gamma$6

with limit set

$\gamma$7

The justifying beliefs are represented by conditional non-standard probability systems, which satisfy probability-on-conditioning-event and the chain rule, and may assign infinitesimal probabilities. The key epistemic operator is c-strong belief, a cautious analogue of strong belief: it keeps all strategies possible, while ranking those in the believed event as infinitely more plausible than those outside it. The paper proves the equivalences

$\gamma$8

and

$\gamma$9

so the belief-based definition coincides both with the standard-priority prudent-rationalizability procedure and with iterated admissibility (Vito, 27 Nov 2025).

Selective Rationalizability introduces a different but related procedure for dynamic games in which players begin with exogenous theories $k$0 about opponents’ behavior. The defining feature is epistemic priority: when observed behavior conflicts with both rationality and the theory, players keep the orders of belief in rationality that remain compatible with the observation and drop the incompatible beliefs in the theory. Formally,

$k$1

and for $k$2, a strategy survives iff there exists $k$3 satisfying sequential best reply, strong belief in previously surviving selectively rationalizable strategies, and strong belief in all rationalizable strategy sets. The epistemic characterization is

$k$4

with limit

$k$5

This model refines Rationalizability, preserves rationality first, and contrasts with Strong-$k$6-Rationalizability, which captures the opposite epistemic priority choice (Catonini, 2017).

These two lines of work describe procedural rational players as cautious, evidence-sensitive, and recursively interpretive. They do not optimize once and for all; they continually refine the set of plausible opponent behaviors in response to observed play.

3. Limited foresight and nested opponent models

A distinct conception of procedural rationality arises in extensive games with limited foresight. The basic game form is

$k$7

and a player at history $k$8 may inspect only a finite, prefix-closed set of continuations via a sight function

$k$9

To evaluate unseen future lines, the model introduces a forked extension $i$0 and heuristic utilities

$i$1

An extensive game equipped with $i$2 and $i$3 is a Monte-Carlo Tree Game, and with an explicit belief structure it becomes an Epistemic Monte-Carlo Tree Game (Turrini, 2015).

Higher-order reasoning is encoded by history-sequences

$i$4

which represent chains of beliefs about what later players can see and how they evaluate what they see. A sight-compatible belief structure $i$5 contains a history-belief map $i$6 and a payoff/evaluation map $i$7, together with correctness of self-sight and monotonicity conditions. The solution concept is the Nested Beliefs Solution, defined recursively over belief-restricted subgames; the global equilibrium notion is the Sight-Compatible Epistemic Solution, under which every move must lead to an NBS outcome in the relevant believed sight. The model collapses to sight-compatible backward induction under coherence conditions and to standard backward induction when beliefs coincide with the actual subgame and actual preferences. The paper also states that for a finite EMTG, computing a SCES is complete for the same complexity class as BI, with algorithmic upper bound

$i$8

and recurrence

$i$9

under the stated size relations (Turrini, 2015).

In this framework, a procedural rational player is neither a full-foresight backward-induction agent nor a merely local short-sight agent. The player is rational relative to what she can inspect and relative to what she believes others can inspect and evaluate.

4. Measurement and identification under belief control

Procedural rationality has also been treated as an empirical object. One approach measures higher-order rationality by pairing human subjects with computer players known to be fully rational, thereby disentangling limited reasoning capacity from belief formation and social biases. In the Robot Treatment, subjects are told that the computer maximizes its own payoff, believes everyone maximizes payoff, and believes everyone believes the computer maximizes payoff. In the History Treatment, subjects face past choice data from human participants in the Robot Treatment, which keeps the observed action distribution while altering the belief environment (Chen et al., 2023).

The design uses 299 NTU students in 41 sessions, two within-subject scenarios, two treatment orders, and two dominance-solvable game families: ring games and guessing games. Revealed rationality is classified by the highest level of iterated dominance consistent with the observed action. The overall treatment-level classification is

$k$0

The central empirical result is that the Robot Treatment distribution of rationality levels stochastically dominates the History Treatment distribution. The reported tests are $k$1 for overall levels, $k$2 within ring games, and $k$3 within guessing games; within-subject, 85% of subjects show weakly higher rationality in the Robot Treatment than in the History Treatment, with Wilcoxon signed-rank test $k$4. Even against robots, around 70% remain below third-order rationality overall (Chen et al., 2023).

The same study tests whether strategic reasoning is stable across games. In the Robot Treatment, about 38.23% of subjects have the same rationality level across ring and guessing games; the simulated mean under independent type draws is 32.80%, with $k$5. In the History Treatment, the corresponding values are 41.30%, 40.27%, and $k$6. Additional metrics support stable relative rankings: the pooled switch ratio in the Robot Treatment is 0.30, the simulated null is approximately 1.01, and $k$7. CRT positively predicts rationality levels, farsightedness also positively predicts rationality levels, and short-term memory does not significantly predict rationality levels (Chen et al., 2023).

This measurement program treats procedural rational players as players whose behavior reflects the process of reasoning rather than merely the outcome of beliefs about others. A plausible implication is that belief control can make strategic capacity more observable than ordinary human-versus-human play does.

5. Counterfactual, structural, and embedded conceptions of rational procedure

One alternative replaces best-response reasoning with perfect prediction. Under Perfect Prediction, players satisfy Necessary Rationality and Necessary Knowledge of Strategies in all possible worlds. The solution concept is the Perfectly Transparent Equilibrium, defined by iterated elimination of non-individually-rational strategy profiles. The first-round survivor set is

$k$8

and later rounds recompute the maximin threshold on the surviving set. The paper proves that, if a PTE exists, it is unique and Pareto-optimal, and on symmetric games it coincides with Hofstadter’s superrationality (Fourny, 2017).

A different line studies pure win-lose coordination games through purely rational principles. A game is represented as

$k$9

and a principle is a nonempty class of protocols. The paper defines FIR, NL, SW, $\gamma$0, IOC, $\gamma$1, IRC, and collective variants such as BCR, COC, and CRC, together with symmetry-based principles ECS, EPS, and ES. It proves, among other comparisons,

$\gamma$2

and establishes that no structural principle can solve a structurally indeterminate game (Goranko et al., 2017).

A third approach embeds players into the environment rather than treating them as special objects. Reflective oracles answer queries $\gamma$3 about the probability that machine $\gamma$4 outputs $\gamma$5, while avoiding diagonalization by randomizing on knife-edge cases. Agents are modeled as oracle machines that choose actions by causal decision theory. In the multi-agent construction, if

$\gamma$6

then the mixed strategy profile $\gamma$7 is a Nash equilibrium iff the oracle is reflective on the relevant threshold queries. The paper also states that every Nash equilibrium can be represented by some reflective oracle (Fallenstein et al., 2015).

These paradigms disagree about the operative rational procedure. PTE emphasizes counterfactual dependence and elimination of impossible worlds; WLC reasoning emphasizes structural principles under no communication or conventions; reflective oracles emphasize embedded agents and self-reference. The commonality is procedural rather than substantive: each theory defines rational players by a rule of reasoning, not solely by a terminal action profile.

6. Algorithmic and distributed-systems applications

Procedural rationality has direct algorithmic use in synthesis and verification. In multi-player weighted reachability games on finite directed graphs, player $\gamma$8 commits to a strategy upfront, and the environment then responds rationally, either as a $\gamma$9-fixed Nash equilibrium or through Pareto-optimal cost tuples. The paper defines rational verification problems NCNV, NCPV, UNCNV, and UNCPV, and synthesis problems CNS, NCNS, CPS, and NCPS. Its main complexity results include: NCPV is in P, UNCPV is in PSPACE, CPS is in PSPACE, NCPS was already known to be EXPTIME-complete, NCNV and UNCNV are in PSPACE, CNS is NP-complete, NCNS is undecidable already with a two-player environment, and with a single environment player NCNS is in EXPTIME and EXPTIME-hard (Bruyère et al., 2024).

In distributed consensus, rational players are modeled as strategic deviators in a Byzantine environment. TRAP addresses rational agreement with correct, Byzantine, and rational players, and proves that consensus is solvable when

$R_i^k(\gamma)$0

Its key mechanism is baiting: accountable consensus yields predecisions, a BFT commit-reveal phase reveals proofs-of-fraud, one coalition member is rewarded for betrayal, and disagreement is resolved. The paper further proves that a baiting strategy is necessary and sufficient to solve rational agreement (Ranchal-Pedrosa et al., 2021).

SNARE adapts this architecture to $R_i^k(\gamma)$1 by separating accountable consensus, which produces a predecision, from one-shot finalization, which produces the actual decision. A central result is that appending a single all-to-all broadcast round with the $R_i^k(\gamma)$2 threshold after predecisions yields $R_i^k(\gamma)$3-$R_i^k(\gamma)$4-robustness for coalitions up to $R_i^k(\gamma)$5 without any deposit; above that, full baiting with deposits under $R_i^k(\gamma)$6 extends tolerance to approximately $R_i^k(\gamma)$7. The paper also states that valid-candidacy holds unconditionally regardless of the quorum threshold, removing both the $R_i^k(\gamma)$8 and $R_i^k(\gamma)$9 constraints from the original TRAP, and that the binding constraint becomes winner consensus on the residual $$R_i^{k+1}(\gamma) \subset R_i^k(\gamma)\quad \forall k \in \mathbb{N}_0$$0 players after exclusion of detected equivocators (Ranchal-Pedrosa et al., 24 Mar 2026).

In these applications, procedural rational players are not merely analytical abstractions. They are explicitly modeled components of systems in which verification, synthesis, consensus safety, and incentive design depend on how rational responses are procedurally generated, constrained, and anticipated.