PaperlayerView exact revision on arXiv

Source-linked AI summary

Entropic Risk-Sensitive Evolutionary Learning and Equilibrium Selection in Coordination Games

Solaleh Mohammadi, Xiang Gao, Kaiqing Zhang

arXiv:2609.08677v1eess.SYcs.GTcs.MAcs.SY

TL;DR

How risk attitudes alter long-run equilibrium selection through evolutionary learning remains comparatively little studied. The paper incorporates entropic risk sensitivity into noisy best responses and analyzes coordination games; it finds that risk attitudes can shift selection between payoff-dominant and maximin outcomes, with super-dominant equilibria robust across risk attitudes.

  • Problem

    How risk sensitivity changes long-run equilibrium selection remains sparsely studied, despite risk-neutral evolutionary learning often selecting the risk-dominant equilibrium.

    Source: §1 · p. 2

  • Method

    The paper incorporates entropic risk sensitivity into noisy best-response learning and analyzes best response with mutations and logit choice, including k-action games.

    Source: §1 · p. 2, §3.1 · p. 6, §6.2 · p. 26

  • Results

    Risk-seeking favors payoff dominance and risk-aversion favors maximin; super-dominant equilibria remain stable across risk attitudes, while sufficiently large k-action populations select strongly dominant outcomes at sufficiently extreme attitudes.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Risk attitudes may provide a systematic way to steer long-run equilibrium selection toward payoff-dominant or maximin outcomes.

    Source: §1 · p. 3

    The analysis assumes homogeneous risk attitudes within each population, uses entropic risk sensitivity, and focuses on coordination-game environments.

    Source: §7 · pp. 33–34

Abstract

from arXiv · show

We study risk-sensitive evolutionary learning dynamics and their long-run equilibrium selection behaviors in coordination games. Agents' risk attitudes enter through the classical entropic risk measure, which evaluates opponent-induced payoff uncertainty and feeds into noisy best responses under two standard revision protocols: best response with mutations and logit choice. We first analyze $2\times 2$ coordination games in both single-population symmetric and two-population asymmetric settings. In the single-population setting, unlike the risk-neutral case where the dynamics are known to favor the risk-dominant equilibrium, we show that risk sensitivity can change the stochastically stable outcome: a greater risk-seeking attitude favors the payoff-dominant equilibrium, while a greater risk-averse attitude favors the maximin equilibrium. Thus, the population's risk attitude may act as a control knob for long-run equilibrium selection. In both population settings, we also identify a robust regime: any super-dominant equilibrium is stochastically stable for all risk attitudes, under both protocols, and across populations. We further extend the single-population analysis to symmetric $k$-action games, which include symmetric $k$-action coordination games as a special case, under risk-sensitive best response with mutations. In this setting, we show that, for sufficiently large populations, sufficiently risk-seeking agents uniquely select the strongly payoff-dominant equilibrium when it exists, whereas sufficiently risk-averse agents uniquely select the strongly maximin equilibrium when it exists. These results show that entropic risk sensitivity may serve as a systematic mechanism for steering equilibrium selection in evolutionary games, beyond the classical risk-neutral benchmark.

1 Introduction

The paper examines whether risk attitudes in evolutionary learning change long-run equilibrium selection. It finds that risk sensitivity can shift selection between payoff-dominant and maximin outcomes, while super-dominant equilibria remain stable across risk attitudes and the mechanism extends to k-action games.

1 Introduction

  • The paper addresses limited understanding of how risk attitudes incorporated into learning dynamics alter equilibrium selection in the original game.Earlier risk-neutral evolutionary learning often uniquely selected the risk-dominant equilibrium.Source: §1 · p. 2
  • The analysis uses the entropic risk measure in noisy best responses under best response with mutations and logit choice, including asymmetric two-population games with heterogeneous risk parameters.The study compares risk-sensitive learning dynamics rather than changing the underlying game’s equilibrium concept.Source: §1 · pp. 2–3
  • In single-population symmetric 2 × 2 games, greater risk-seeking favors payoff dominance and greater risk-aversion favors maximin when those equilibria differ, under both protocols.A super-dominant equilibrium is stochastically stable for all risk attitudes under both protocols and in both population settings.Source: §1 · p. 2
  • In single-population symmetric k-action games under risk-sensitive BRM, sufficiently large populations select the strongly payoff-dominant equilibrium under sufficient risk-seeking, or the strongly maximin equilibrium under sufficient risk-aversion, when each exists.The paper presents risk attitudes as a way to steer long-run equilibrium selection.Source: §1 · p. 3

2 Preliminaries

The preliminaries define the coordination-game setting and the equilibrium refinements used to compare outcomes. They also specify the one- and two-population evolutionary setup and how payoffs depend on population states.

2 Preliminaries

  • A 2 × 2 coordination game has two strict Nash equilibria, the coordinated profiles (1, 1) and (2, 2).In the symmetric case, the coordination conditions reduce to a > c and d > b.Source: §2 · p. 3
  • Risk dominance compares the products of deviation-payoff losses, reducing in the symmetric setting to a − c > d − b.The reverse inequality makes (2, 2) risk-dominant.Source: §2 · p. 3
  • Payoff dominance means one coordinated equilibrium Pareto-dominates the other, while maximin favors an action with the higher worst-case payoff.A super-dominant equilibrium is weakly maximin for both agents and weakly payoff-dominant, with at least one condition strict.Source: §2 · p. 4
  • The model allows one or two populations choosing between actions 1 and 2, with matching within the same population in the single-population setting and across populations otherwise.Payoffs depend on the opponent population’s state in the two-population setting.Source: §2 · pp. 4–5

3 Risk-Sensitive Evolutionary Dynamics & Stochastic Stability

The section defines risk-sensitive noisy best responses and the stochastic-stability framework used to analyze them. It specifies the BRM and logit protocols, then connects small-noise transition costs to stable outcomes.

3.1 Risk-Sensitive Evolutionary Dynamics

  • The entropic risk measure replaces expected-payoff evaluations with risk-sensitive evaluations of opponent-induced payoff uncertainty; positive β favors favorable realizations and negative β emphasizes adverse ones.At β = 0, the measure reduces to expected payoff.Source: §3.1 · p. 5

3.1.1 Best Response with Mutations (BRM)

  • Under BRM, agents choose a strict ERM-adjusted best response with probability 1 − e^-1/η, choose the alternative with probability e^-1/η, and break ties uniformly.The resulting single- and two-population dynamics are called RS-BRM and 2Pop-RS-BRM.Source: §3.1.1 · p. 7

3.1.2 Logit Choice

  • Under logit choice, actions with higher ERM-adjusted payoffs receive higher choice probability, and the rule converges to an ERM-adjusted best response as noise vanishes.The resulting dynamics are called RS-logit and 2Pop-RS-logit.Source: §3.1.2 · p. 7

3.2 A Unified Framework for Stochastic-Stability Analysis

  • For fixed population size, the noisy dynamics form an irreducible, aperiodic finite-state Markov chain with a unique stationary distribution, and stochastic stability concerns its small-noise limit.Transition costs capture the exponential decay of one-step probabilities; zero-cost transitions follow ERM-adjusted best responses or ties.Source: §3.2 · pp. 8–9
  • Under the framework’s two-attracting-state assumption, stochastic-stability analysis reduces to comparing the minimal costs of escaping the two stability sets.The zero-noise dynamics describe limiting motion as η → 0.Source: §3.2 · p. 10

4 Single-Population Symmetric Setting

In symmetric 2 × 2 coordination games, risk attitudes can shift long-run selection between maximin and payoff-dominant equilibria, while super-dominance makes selection robust to risk attitudes under both protocols.

4 Single-Population Symmetric Setting

  • The large-population conclusions are pointwise in risk attitude, and the required population threshold need not remain bounded near a critical value.The threshold may diverge as risk sensitivity approaches the cutoff.Source: §4 · p. 15
  • A super-dominant equilibrium is uniquely stochastically stable for every risk attitude under both protocols when the population is sufficiently large.Super-dominance aligns payoff dominance with protection against unfavorable mismatches.Source: §4 · pp. 15–16
  • The two noisy best-response protocols produce the same equilibrium-selection pattern, despite using different transition-cost rules.The analysis uses escape costs and a selection coefficient to determine which attracting state is uniquely stable.Source: §4 · pp. 12–15
  • When payoff-dominant and maximin equilibria differ, sufficiently risk-seeking agents select the payoff-dominant equilibrium, whereas sufficiently risk-averse agents select the maximin equilibrium under both protocols.For each fixed risk attitude, these results hold for sufficiently large populations; the cutoff can differ between protocols.Source: §4 · pp. 12, 14, 16

5 Extension: Two-Population Asymmetric Setting

In asymmetric two-population games, stochastic selection under best response with mutations depends on relative escape costs, which vary with both populations’ risk attitudes. The analysis identifies the zero-noise attracting states and characterizes when super-dominance guarantees selection for sufficiently large populations.

5 Extension: Two-Population Asymmetric Setting

  • The two-population model permits distinct risk parameters across populations, while each population is homogeneous in its risk attitude.Unlike the single-population model, its aggregate state is two-dimensional and each population’s best response depends on the other population’s state.Source: §5 · p. 17
  • The general risk-sensitive two-population logit selection problem remains open because transition costs are state- and path-dependent.The analysis characterizes stochastic selection under super-dominance, while best response with mutations remains tractable in the two-dimensional setting.Source: §5 · p. 17

1. Each such move starts from a state in BN 1,N 2

  • The zero-noise dynamics have exactly two attracting states, (0, 0) and (1, 1), with lower-left and upper-right rectangular stability sets.Every other state has zero-cost paths to both attracting states.Source: §5 · pp. 18–19
  • Under two-population risk-sensitive best response with mutations, the stable equilibrium is determined by the relative escape costs, which depend on both populations’ risk parameters.Changing risk attitudes can alter the relative difficulty of escaping the coordinated equilibria and thereby change which one is stochastically stable.Source: §5 · p. 20
  • For sufficiently large populations, a super-dominant equilibrium is uniquely stochastically stable under best response with mutations for every fixed pair of risk parameters.The analogous claim holds for either coordinated equilibrium when it is super-dominant.Source: §5 · p. 20

1. Each such move starts from a state in BN 1,N 2

In the two-population setting, best-response-with-mutations selection is governed by risk-dependent escape costs, but super-dominance makes selection robust to heterogeneous risk attitudes for sufficiently large populations. A numerical illustration shows finite-time coordination shifting with one population’s risk attitude, with the transition location also depending on the other population’s attitude.

1. Each such move starts from a state in BN 1,N 2

  • Under two-population best response with mutations, equilibrium selection is determined by the relative escape costs, which depend on the risk-sensitive thresholds and hence on both risk parameters.Changing risk attitudes can alter escape difficulty and the stochastically stable equilibrium.Source: §1 · p. 21
  • A super-dominant equilibrium remains uniquely stochastically stable under both revision protocols for fixed risk-attitude profiles when both populations are sufficiently large.This robustness extends the single-population result and holds even when the populations have heterogeneous risk preferences.Source: §1 · pp. 21–22

5.1 Numerical Illustration

  • As β1 increases, more simulated runs end near (1, 1); as β1 decreases, more end near (0, 0), and the transition location depends on β2.The finite-time coordination frequencies transition near the interval where the two escape costs tie.Source: Figure 2 · p. 23, §5.1 · p. 24
  • The numerical illustration measures terminal coordination frequencies and finite-population escape costs as β1 varies for three fixed values of β2.The frequency panel includes pointwise 95% Wilson intervals; the escape-cost panel shades equal-cost regions.Source: Figure 2 · p. 23, §5.1 · pp. 23–24

6 Extension: k-Action Setting

This section extends single-population risk-sensitive evolutionary analysis to symmetric games with finitely many actions, examining equilibrium selection under best response with mutations. Its analysis and numerical illustration show how risk attitudes can shift selection toward strongly payoff-dominant or strongly maximin equilibria.

6.1 Model and Setup

  • In k-action coordination games, payoff dominance is equivalent to strong payoff dominance, while strong maximin is stricter than strict maximin.Strong payoff dominance compares an action's diagonal payoff with payoffs from alternatives against any opponent action; strong maximin compares its worst-case payoff with alternatives' best payoffs against it.Source: §6.1 · pp. 25–26

6.2 Risk-Sensitive Evolutionary Dynamics

  • The k-action extension studies a single-population best-response-with-mutations protocol in which revisions use entropic-risk-adjusted best responses.In the vanishing-noise limit, revisions to adjusted best responses have zero cost, while revisions to other actions have unit cost.Source: §6.2 · pp. 26–27

6.3 Analyses

  • The analysis uses a radius–coradius criterion: an attracting state is uniquely stochastically stable when its radius exceeds its coradius.The criterion is stated for the finite-state RS-BRM Markov chain.Source: §6.3 · p. 28
  • For sufficiently large populations, RS-BRM uniquely selects a strongly payoff-dominant equilibrium under sufficiently risk-seeking behavior, or a strongly maximin equilibrium under sufficiently risk-averse behavior.The corresponding risk-sensitivity thresholds depend on the equilibrium; the result applies to finite symmetric k-action games.Source: §6.3 · pp. 28–29

6.4 Numerical Illustration

  • A four-action example illustrates the two selection effects: one diagonal equilibrium is strongly payoff-dominant and another is strongly maximin.The example is not a four-action coordination game because (1, 1) is not a strict Nash equilibrium.Source: §6.4 · p. 31
  • For β < β4, Theorem 11 predicts that e4 is the unique stochastically stable state for all sufficiently large N.Source: §6.4 · p. 32
  • As β increases above zero, the high diagonal payoff a22 = 10 makes action 2 increasingly attractive, selecting the strongly payoff-dominant equilibrium (2, 2).Positive β emphasizes high-payoff realizations.Source: §6.4 · p. 33
  • For β < 0, action 4 guarantees at least 2 against every opponent action, while other actions yield at most 0 against action 4, favoring strongly maximin (4, 4).Negative β emphasizes low-payoff realizations and downside protection.Source: §6.4 · p. 33

7 Concluding Remarks

The paper concludes that entropic risk sensitivity can shift long-run equilibrium selection in coordination games, while super-dominant equilibria remain stable across risk attitudes and protocols. It also extends the selection results to symmetric k-action games and identifies limits of the analysis.

7 Concluding Remarks

  • In single-population symmetric 2 × 2 coordination games, risk-seeking favors the payoff-dominant equilibrium and risk-aversion favors the maximin equilibrium when they differ, under both protocols.The maximin equilibrium is described as worst-case robust.Source: §7 · p. 33
  • Super-dominant equilibria remain stochastically stable for all risk-sensitivity parameters in both single- and two-population settings.When payoff dominance and worst-case robustness align, selection is insensitive to risk attitudes and the noisy best-response protocol.Source: §7 · p. 33
  • In symmetric k-action games, sufficiently risk-seeking agents select strongly payoff-dominant equilibria, while sufficiently risk-averse agents select strongly maximin equilibria, when those equilibria exist.The conclusion applies to the extended single-population analysis.Source: §7 · p. 33
  • The analysis assumes homogeneous risk attitudes within each population and focuses on coordination-game environments.The paper identifies heterogeneous or time-varying risk sensitivity, other risk-sensitive criteria, richer population structures, and other payoff structures as directions for further study.Source: §7 · pp. 33–34

A.1 Proof of Lemma 1

The proof establishes that the relevant function is strictly increasing, giving a unique threshold root in the coordination-game setting. Its sign characterization follows from that monotonicity.

A.1 Proof of Lemma 1

  • The derivative argument uses the coordination-game inequalities a > c and d > b.Source: §A.1 · p. 34
  • The proof establishes that ∆′(z) > 0 throughout (0, 1), so ∆ is strictly increasing there.The derivative argument accounts for the reversed sign when β < 0 by the negative factor 1/β.Source: §A.1 · p. 34
  • Strict monotonicity and continuity give a unique root of ∆ in (0, 1).The proof derives the root for β ≠ 0 and obtains the β = 0 expression by continuous extension or by solving ∆₀(z) = 0.Source: §A.1 · p. 35

A.2 Proof of Lemma 2

The proof analyzes how the comparison function F varies with risk sensitivity and uses its monotonicity and limiting values to characterize the threshold behavior. It also treats cases where intervals defined by payoff differences are nested.

A.2 Proof of Lemma 2

  • F is monotone on all of R, with its behavior derived from a derivative argument and continuity at β = 0.The proof states that F extends continuously to β = 0, where F(0) = (b + d − a − c)/2.Source: §A.2 · pp. 35–36
  • The endpoint limits of F are d − a as β → ∞ and b − c as β → −∞, which determine the stated monotonicity cases for x*β.When both endpoint limits are positive, F(β) > 0 for all β; when both are negative, F(β) < 0 for all β.Source: §A.2 · p. 36
  • When the intervals [b, d] and [c, a] are properly nested, the proof establishes whether x*β stays below or above 1/2 across nonzero β.The cases use [b, d] as a proper subinterval of [c, a], or the reverse, and separately check β = 0.Source: §A.2 · p. 36

A.3 Proof of Theorem 3

The proof compares minimum-cost rooted trees and escape costs to establish stochastic stability, with Assumption 2 restricting the zero-noise dynamics to recurrent classes {A} and {B}.

A.3 Proof of Theorem 3

  • The finite rooted-tree sums defining stationary weights are exponentially dominated by their minimum-cost trees, whose costs define stochastic potentials.Each tree is rooted at a state and gives every other state a unique directed path to that root.Source: §A.3 · p. 37
  • The proof identifies the stochastic potentials of A and B with their corresponding escape costs.For the upper bound, it constructs an A-rooted spanning tree using a path from B to A and zero-cost paths for the remaining states; the lower bound uses the path from B that exits its stability set.Source: §A.3 · pp. 37–38
  • The zero-noise transition matrix is the entrywise limit as η → 0, with positive-probability off-diagonal transitions exactly matching zero-cost transitions.Source: §A.3 · p. 38
  • Under Assumption 2, the only recurrent classes of P_N,0 are the singleton sets {A} and {B}.Any other recurrent class would contain a state with a zero-cost path to A or B, contradicting closedness.Source: §A.3 · p. 38
  • The limiting stationary mass satisfies µ_N,η(A) → 1 and µ_N,η(B) → 0, so A is uniquely stochastically stable.The proof takes subsequential limits of stationary distributions and uses that limiting stationary distributions are supported on recurrent classes.Source: §A.3 · pp. 38–39
  • A separate conclusion in the proof states that B is uniquely stochastically stable.Source: §A.3 · p. 39

Appendix B Auxiliary Lemmas

The auxiliary lemmas establish continuity and monotonicity properties of the logit function and a positivity inequality across risk attitudes. They also identify the risk-neutral case with expected-payoff evaluation.

Appendix B Auxiliary Lemmas

  • The function S̄_logit is continuous and strictly monotone unless |a − b| = |c − d|, when it is constant.The stated properties hold on R.Source: Appendix B · pp. 39–40
  • When βp = 0, the entropic-risk-measure payoffs reduce to expected payoffs.Source: Appendix B · p. 41
  • R_βp(z) is positive for every βp ∈ R and z ∈ [0, 1].The proof establishes positivity using R_0(z) = (a_p + b_p) − (c_p + d_p) > 0.Source: Appendix B · p. 41
Loading 2609.08677v1…