acceptodds
Under review as a conference paper at ICLR 2027

GraphZD: Graph-Aware Continuous Zero-Determinant Control in Evolutionary Networks

Abstract

The emergence and maintenance of cooperation in finite, structured populations remain a fundamental challenge in evolutionary game theory. Zero-determinant (ZD) strategies can enforce linear relations between long-term payoffs, yet fixed extortionate or generous control cannot adapt to heterogeneous neighborhood conditions. This limitation is particularly important in evolutionary networks, where local structure shapes both interaction payoffs and strategy replacement. In this paper, we propose GraphZD, a graph-aware continuous ZD controller for repeated Prisoner's Dilemma games on evolutionary networks. The method instantiates a graph-aware continuous relational-control paradigm that treats the enforceable payoff relation, rather than a predefined strategy label, as the object of adaptation. A graph attention encoder summarizes local behavior, payoff, and structure. An online contextual linear controller evaluates nine control anchors and forms a continuous proposal, which is regulated by an identification phase, exponential moving-average smoothing, and a coexistence gate. An analytically computed common scale then maps the control to a feasible memory-one multiplicative ZD policy without coordinate-wise clipping, preserving the intended Press-Dyson vector direction. We evaluate GraphZD over 5,000 death-birth generations on grid, scale-free, SBM, and small-world networks initialized with 20% adaptive carriers and 80% AllD. In paired experiments, GraphZD improves population share, cooperation, and collective payoff over an otherwise identical controller without graph information. It also outperforms fixed extortionate and generous ZD endpoints: relative to the stronger generous endpoint, GraphZD raises mean population share by 17.8 percentage points and mid-run joint payoff by 0.84. The trajectories support a two-stage interpretation: graph-aware control improves survival under early defector pressure, allowing cooperative groups to persist and generate higher social welfare later. These results connect local structural information with adaptive payoff control in finite evolutionary populations.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural networks, Evolutionary game theory, Zero-determinant strategies, Contextual bandits, Structured populations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4BPEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GraphZD, a graph-aware continuous zero-determinant (ZD) controller for evolutionary networks.

The method uses a graph attention network to encode local behavior, payoffs, and network structure. A contextual linear controller then selects continuous ZD control parameters based on local feedback. The authors also introduce a feasible mapping that preserves the Press–Dyson direction without independently clipping cooperation probabilities.

Experiments on multiple network structures show improvements in population share, cooperation, and social welfare compared with a controller without graph information and fixed ZD strategies.

Critical strengths
  1. A meaningful combination of graph information and adaptive ZD control.

The paper addresses an important limitation of fixed ZD strategies in structured populations. The proposed framework adapts payoff control based on local network conditions. The feasible mapping is mathematically justified and preserves the intended ZD direction. This provides a sound foundation for the proposed method.

  1. Careful experimental design with consistent improvements.

The authors evaluate GraphZD across several network structures using paired experiments and statistical tests. Compared with the controller without GNN, GraphZD improves mean population share by 2.9 percentage points, mid-run cooperation by 11.8 percentage points, and mid-run social welfare by 0.451.

The additional experiments on network size, receptive fields, and mixed populations provide further evidence of the method's effectiveness.

Critical weaknesses
  1. The necessity of learned continuous adaptation is not fully established.

The paper compares GraphZD with a controller without graph information and two fixed ZD endpoints. These comparisons are useful, but they do not fully isolate the benefit of learning continuous control parameters.

In particular, the paper lacks comparisons with a fixed fair ZD strategy and simpler adaptive strategies based on local neighborhood statistics. Moreover, the fixed endpoints do not use the same smoothing and gating mechanisms.

These comparisons are important to determine whether the improvements come from the proposed adaptive controller rather than simpler control choices.

  1. The proposed two-stage evolutionary mechanism lacks direct validation.

The authors suggest that graph-aware control supports early survival against defectors and later cooperative consolidation. The population and welfare trajectories are consistent with this explanation.

However, the paper does not directly show how the learned control parameter changes over time or across different neighborhood conditions. It is therefore unclear whether GraphZD actually follows the proposed transition from extortion-oriented to generosity-oriented behavior.

  1. The connection between theoretical ZD control and finite-game experiments needs further examination.

The paper proves that its mapping produces feasible cooperation probabilities and preserves the Press–Dyson direction. However, the payoff relation is defined for long-run interactions, while the experiments use 20-round repeated games with memory reset between generations.

The paper does not directly measure how closely the realized payoffs satisfy the intended ZD relation under this finite interaction setting. This limits the empirical validation of the proposed payoff-control mechanism.

Questions for authors
  1. Could the authors compare GraphZD with a fixed fair ZD strategy and simple neighborhood-based adaptive policies? This would help isolate the benefit of learned continuous control.

  2. Could the authors visualize the distribution of the control parameter over generations? In particular, does it become more extortion-oriented near defectors and more generosity-oriented as cooperative groups grow?

  3. How closely do the observed payoffs satisfy the intended ZD relation during the 20-round interactions? Does increasing the number of rounds change the results?

Reviewer_btHBOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses how ZD strategies can adapt to changing neighborhood conditions in networked evolutionary games. GraphZD combines graph representations with online continuous control to adjust between extortion and generosity while preserving ZD feasibility. Experiments across several network topologies show improved invasion, cooperation, and social welfare compared with fixed ZD strategies and a non-GNN variant.

Critical strengths

1 The paper addresses a meaningful trade-off between resisting exploitation and sustaining cooperation, using neighborhood context to adapt within a feasible ZD strategy family. This provides an interpretable approach to evolutionary strategy adaptation. 2 Evaluation across multiple network topologies, population sizes, and repeated runs provides evidence that the observed improvements extend beyond a single experimental setting.

Critical weaknesses

1 The benefits of continuous, graph-aware control are not sufficiently isolated. PERG is evaluated only in exploratory mixed-population competition, while simple neighborhood-statistics controllers and a fixed fair ZD baseline are missing. Consequently, the results do not establish whether the proposed complexity is necessary or outperforms simpler adaptive alternatives. 2 Key design choices lack sufficient justification and validation. The feedback signals, identification phase, EMA, gating, and fair fallback are not adequately isolated through ablations. Moreover, the long-run ZD payoff relation is not assessed under the 20-round, memory-reset protocol, and sensitivity to the fitness offset remains unclear. These gaps limit confidence in the mechanism behind the gains and their robustness.

Questions for authors

1 Baseline comparison: How does GraphZD compare with PERG under the same invasion protocol, a controller using simple neighborhood statistics, and a fixed fair ZD strategy (z=0z=0)? These comparisons would help establish the contribution of continuous control and learned graph representations. 2 Component contribution: What ablation evidence identifies the contributions of the three feedback signals, identification phase, EMA, coexistence gate, and fair fallback? Which components are essential to the reported gains? 3 Finite-horizon validity: How closely do realized payoffs satisfy the intended ZD relation under 20-round interactions with memory reset each generation? Are the main findings robust to changing the interaction length? 4 Selection sensitivity: Are the reported advantages robust to changes in the fitness offset of 0.1, which affects relative reproductive probabilities? 5 Seed selection clarification: What does “previously screened” mean for the five seeds included in the mixed-strategy experiment? Were any seeds excluded based on outcomes, and do the conclusions hold when those five seeds are omitted?

Reviewer_9gLJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes GraphZD, a method that adjusts zero-determinant (ZD) strategies using information about a player’s neighborhood. A frozen GNN summarizes local behavior, payoffs and network structure. These features feed into an online linear controller, which learns from feedback and scores nine control settings using a UCB exploration bonus. It combines these settings into a continuous choice between more extortionate and more generous behavior, then smooths and maps that choice to valid ZD cooperation probabilities.

The authors test GraphZD in repeated Prisoner’s Dilemma games on grid, scale-free, small-world and community-structured networks, including different population sizes and a mixed-strategy competition. In general, they show that GraphZD spreads faster and achieves higher intermediate cooperation and welfare than the version without the GNN and fixed ZD strategies. Additionally, their algorithm has an advantage over the non-GNN versio, particularly in larger populations, although final outcomes sometimes become similar. GraphZD also remains competitive with other adaptive strategies in the mixed-population grid experiment.

Critical strengths
  1. The experiments show a consistent benefit across several settings.
    The paper considers different network types and population sizes, reports uncertainty across runs, and compares one-hop and two-hop variants. The population-size results are also interesting, where the advantage over the version without the GNN becomes much larger in bigger populations. In several experiments, GraphZD also has lower variability across runs. Together, these results provide useful evidence that the method can improve both performance and consistency.

  2. The paper includes useful checks beyond the main performance curves.
    The reward-prediction results suggest that the GNN features help the controller predict welfare, payoff advantage and neighborhood retention. This gives a possible explanation for the observed gains. The feasibility proof and numerical audit also check that the controller produces valid cooperation probabilities. Notably, the authors acknowledge the limited scope of their theoretical analysis and describe the mixed-population comparison as exploratory.

Critical weaknesses
  1. The experiments do not yet show whether a GNN is necessary.
    The main comparison is against a version “without GNN.” If this also removes neighborhood information, the result mainly shows that local information helps. It does not show that a GNN is better than a simpler way of using that information. For example, the same controller could receive neighbors’ average cooperation, payoff means and variances, the player’s relative payoff, and degree statistics. These are easy to interpret and could capture much of the useful context. The prediction and population-size results strengthen the case for the current representation, but a comparison with these simple features is still needed to justify its complexity.

  2. The broader contribution over existing adaptive strategies is not fully established.
    PERG [Huang et al., 2026] already switches between extortion and generosity based on the two players’ running payoffs. GraphZD adds learned contextual adaptation and continuous control, so there is a real difference. However, PERG is tested only in one mixed-population grid experiment, while most other experiments compare internal variants or fixed ZD strategies. That makes it difficult to judge how much the full method improves over existing adaptive approaches. The theory does not resolve this either: the regret result concerns a simplified hard-UCB controller, rather than the implemented combination of averaging, gating and smoothing.

Questions for authors
  1. Could you compare the GNN with simple neighborhood features while keeping the rest of the controller unchanged?
    Useful features could include cooperation and payoff means and variances, the player’s payoff relative to its neighbors, the fraction of neighbors earning more, and basic degree statistics. Relative-payoff features are also motivated by PERG’s switching rule. Including this baseline in both the main experiments and the population-size study would help separate the value of neighborhood information from the value of GNN encoding.

  2. Could you strengthen the comparison with PERG under the main experimental protocol?
    Can the authors show how PERG performs in the main experiments where 20% of players use the tested strategy and the remaining 80% always defect? Using the same network types and population sizes would make the comparison with GraphZD clearer. Can the authors also explain whether PERG’s running payoff averages are reset between interactions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.