Axelrod-Style Tournament

• Axelrod-style tournaments are round-robin competitions where Iterated Prisoner’s Dilemma strategies are evaluated by aggregated payoffs to study cooperation, retaliation, and adaptation.
• They evolved from Axelrod’s historical studies to modern frameworks like Axelrod-Python, incorporating fixed match lengths, noise elements, and extensive reproducibility protocols.
• Key insights reveal that high cooperation and strategic flexibility, as exemplified by Tit for Tat, are crucial for success in diverse and noisy competitive environments.

Searching arXiv for the specified paper and closely related work on Axelrod-style tournaments, reproducibility, and IPD strategy evaluation. An Axelrod-style tournament is a round-robin competition among computer programs or other formalized strategies that play the Iterated Prisoner’s Dilemma (IPD), with ranking determined by payoffs aggregated across the opponent pool rather than by knockout elimination or isolated head-to-head success (Knight et al., 17 Oct 2025). The format was pioneered by Robert Axelrod in the early 1980s to study direct reciprocity, and it became a canonical experimental paradigm because it connected a simple repeated game to computational questions about cooperation, retaliation, forgiveness, stochasticity, and adaptation (Knight et al., 2016). In contemporary usage, the term encompasses both historical reproductions of Axelrod’s original events and larger open tournaments implemented in frameworks such as Axelrod-Python, often with extensions for noise, probabilistic termination, and evolutionary or ecological follow-on analyses (Knight et al., 17 Oct 2025).

1. Definition and formal setting

The underlying one-shot game is the Prisoner’s Dilemma, a $2\times 2$ interaction in which each player chooses Cooperate ($C$) or Defect ($D$). Its payoffs satisfy the standard inequalities $T > R > P > S$ and $2R > T + S$, where $T$ is Temptation to defect, $R$ is Reward for mutual cooperation, $P$ is Punishment for mutual defection, and $S$ is the Sucker’s payoff. The canonical values used in Axelrod’s original studies and in much of the later literature are $T=5$, $C$0, $C$1, and $C$2 (Knight et al., 17 Oct 2025).

In the IPD, this stage game is repeated over many rounds, and each player may condition its next move on some function of the interaction history. In the simplest analytical treatment, a match of $C$3 rounds yields player $C$4 an average score

$C$5

where $C$6 depends on the joint action at round $C$7. A tournament then aggregates such match scores across all opponents and repeats. The central methodological feature is therefore population-level evaluation: a strategy is not judged solely by whether it can exploit a particular opponent, but by its average performance in a field (Knight et al., 17 Oct 2025).

This tournament perspective differs from equilibrium analysis of infinitely repeated games, although the two are related. The IPD literature also studies discounted infinite interactions, typically with

$C$8

and many memory-one analyses represent pairwise play as a Markov chain over the four outcomes $C$9 (Knight et al., 2016). Axelrod-style tournaments operationalize these abstractions in computational form: they instantiate a finite or probabilistically terminated repeated game, a specified opponent set, and a ranking criterion.

2. Historical tournaments and archival record

Axelrod’s original tournaments established the template. In the first tournament, programs played every other program, a copy of themselves, and a random baseline that cooperated with probability $D$0; matches had fixed length, and ranking was by average payoff across all matches. In the second tournament, the principal change was removal of a known fixed horizon: Axelrod stated that the per-move hazard of termination was $D$1, chosen so the expected median length would be 200 moves (Knight et al., 17 Oct 2025).

Feature	First tournament	Second tournament
Entries	14 entries	63 entries
Match horizon	Fixed length	No fixed length known in advance
Special opponents	Self-play and random baseline $D$2	Surviving Fortran strategies against one another in the reproduction

Both tournaments used the same canonical payoff values $D$3, $D$4, $D$5, $D$6, and in both tournaments Tit for Tat (TFT) emerged as the winner (Knight et al., 17 Oct 2025). The historical influence of these events lies in the qualitative conclusion that simple contingent cooperation could outperform more elaborate designs, and in the later distillation of high-performing properties such as being “nice,” retaliatory, forgiving, and clear (Knight et al., 2016).

The archival record is incomplete. For the first tournament, only the published report survives and the original code is lost. For the second, the submitted Fortran strategies survive online, but Axelrod’s final tournament orchestration code does not (Knight et al., 17 Oct 2025). Related literature summarizes the 1979–80 tournaments as 13 strategies in the first and 64 in the second (Knight et al., 2016). This suggests that historical counting conventions are not uniform, which is one reason the exact reconstruction of the original events remains nontrivial.

The second tournament is especially important because the archival asymmetry is narrower: the individual strategy sources exist, but the control logic, seed choices, and final orchestration do not. That combination makes the event reproducible only in the weaker sense of systematic reconstruction rather than bit-for-bit replay (Knight et al., 17 Oct 2025).

3. Rules, implementations, and scoring conventions

In the restored second-tournament setting, the 63 submitted strategies are Fortran functions with a standardized signature receiving the match state and returning the next action, encoded as $D$7 cooperate and $D$8 defect. Inputs included the opponent’s last move ($D$9), current turn number ($T > R > P > S$0), cumulative scores ($T > R > P > S$1), a random number $T > R > P > S$2 for stochastic decisions, and own last move ($T > R > P > S$3) (Knight et al., 17 Oct 2025). Strategies are now identifiable by their Fortran function names; for example, TFT corresponds to k92r.

A distinctive rule of the second tournament concerns termination. Axelrod described a geometric stopping rule with hazard $T > R > P > S$4, for which the median length under geometric termination is approximately

$T > R > P > S$5

With $T > R > P > S$6, this gives roughly 200 moves. In practice, however, the tournament did not dynamically terminate each match independently. Instead, five match lengths were pre-sampled once and then used identically for every pair of players; the inferred realized set is $T > R > P > S$7, with average length 151 (Knight et al., 17 Oct 2025).

The original Fortran strategies assume an initial cooperative move, and modern adapters preserve this convention, so both players start by cooperating (Knight et al., 17 Oct 2025). This is a technically consequential convention, because first-move behavior affects early-state occupancy, transient punishment cycles, and the interpretation of “nice” strategies. More generally, initial-move conventions are strategy-specific in the broader Axelrod library, which is one reason modern tournament reports are expected to make them explicit (Knight et al., 2016).

Scoring is by average payoff per round within a match and then by aggregation across opponents and repeats. In the historical reproduction, each pairing plays five matches, with the same five fixed lengths used for all pairings (Knight et al., 17 Oct 2025). In later large-scale studies, additional summary statistics have been used, including average payoff per turn, counts of head-to-head match wins, and normalized rank $T > R > P > S$8, where $T > R > P > S$9 denotes the tournament winner (Knight et al., 2016, Glynatsi et al., 2020). The choice of ranking statistic matters: some strategies win many individual matches but still underperform on overall average score.

4. Reproduction and open computational frameworks

The modern reconstruction of Axelrod’s second tournament proceeded by restoring the raw Fortran source from an HTML-embedded file (TourExec1.1.f), stripping formatting, and making minimal changes needed for modern compilation. Each strategy was split into its own file and compiled into a shared library, libstrategies.so, via a dedicated Makefile; a typical install location is /usr/local/lib on POSIX systems (Knight et al., 17 Oct 2025). A thin Python adapter then calls the unmodified Fortran strategy functions from within the Axelrod-Python ecosystem, reconstructing the match-state inputs expected by the legacy code and enforcing the original initial-cooperation convention.

Axelrod-Python provides the tournament engine, match management, aggregation, and reporting. Its design goal is reproducible game-theoretic research into the IPD, with support for strategy design, tournament execution, noisy interactions, probabilistic endings, and ecological simulations such as Moran processes (Knight et al., 2016). In the library’s interface, a tournament is created by instantiating axelrod.Tournament with a list of strategy instances and optional configuration such as noise, match length, or probabilistic termination.

Reproducibility is treated as a software and data-management problem as much as a game-theoretic one. The library paper emphasizes open-source development, standardized strategy interfaces, documentation, version pinning through Zenodo, and extensive testing, reported as approximately $2R > T + S$0 coverage at the time of writing (Knight et al., 2016). The restoration of the second tournament extends this ethos to historical code preservation: the surviving Fortran strategies are made easy to use within a contemporary framework, while the absence of the original orchestration code and historical seeds is handled by repeated independent realizations rather than by claims of exact replay (Knight et al., 17 Oct 2025).

For the reproduced second tournament, the unknown original seed is addressed by running 25,000 independent tournament realizations and averaging out stochastic variation (Knight et al., 17 Oct 2025). This is methodologically consistent with later large-tournament studies, which likewise emphasize repeated trials, fixed seeds where possible, and explicit recording of parameters and versions (Knight et al., 2016).

5. Strategic regularities and tournament outcomes

The reproduction confirms Axelrod’s main qualitative conclusions. TFT (k92r) again ranks first, with average score 2.878 and average cooperation rate 0.922, while the worst performer (k36r) is again at the bottom (Knight et al., 17 Oct 2025). Successful strategies are highly cooperative, retaliatory, and forgiving: they reciprocate cooperation, punish defection promptly, and return to cooperation quickly. Among the top strategies, cooperation against one another is near-symmetric.

At the same time, the historical result is not perfectly deterministic. Across 25,000 reproduced tournaments, TFT wins $2R > T + S$1 of realizations, while k42r wins $2R > T + S$2, implying that $2R > T + S$3 of seeds produce a winner other than TFT (Knight et al., 17 Oct 2025). Rank shifts are usually modest, often 1–3 places, but some are larger. The most notable discrepancy concerns “Champion” (k61r), which falls from original rank 2 to reproduced rank 12, with average score 2.791 and cooperation rate 0.954; the authors identified and corrected a bug in the Fortran code, namely ICOOP not being initialized across matches, which likely explains part of the difference (Knight et al., 17 Oct 2025).

Later large-scale tournament studies complicate the older intuition that a single simple rule can dominate universally. Across 195 strategies and 11,420 unique trials per tournament type, no single strategy performs well in diverse IPD scenarios (Glynatsi et al., 2020). Instead, high performance depends on the environment. The paper refines Axelrod’s older characterization into a set of properties: be nice, be provocable and generous, be a little envious, be clever, and adapt to the environment. One especially strong empirical regularity is that winners often have cooperation rate $2R > T + S$4, meaning that effective performance is closely tied to matching the aggregate cooperation level of the tournament population (Glynatsi et al., 2020).

This broader evidence also clarifies the status of zero-determinant (ZD) strategies. In the Axelrod library paper, ZD extortionate strategies win many individual matches but are not top-ranked by overall average score, a pattern that persists under noise (Knight et al., 2016). In large, heterogeneous pools, pure extortion is therefore not a general route to victory. A related study reports that extortioners extract high payoffs from unconditional cooperators yet perform poorly against retaliatory or non-evolutionary opponents, and extortion versus extortion collapses to mutual defection (Malik, 2021). The population-level ranking criterion of an Axelrod-style tournament therefore penalizes strategies that succeed only pairwise or only against compliant opponents.

Modern tournament champions often use richer representations. Reported top performers include Evolved HMM 5, Evolved FSM 16, EvolvedLookerUp2_2_2, PSO Gambler variants, and Evolved ANN (Glynatsi et al., 2020). In another study, a two-phase strategy combining a noise-aware forgiving TFT variant (“Longterm TFT”) with an adaptive memory-1 optimizer (“ISO”) outperforms all previously existing strategies in round-robin evaluation against 239 Axelrod-library opponents at noise levels $2R > T + S$5, $2R > T + S$6, $2R > T + S$7, and $2R > T + S$8 (Hutter, 2023). A separate line of work evolves an FSM that reaches first place in the full Axelrod tournament, with the final reduced controller being a six-state FSM derived from an eight-state evolved machine (Vincent et al., 2021). These results indicate that the tournament format has become a benchmark not only for hand-designed reciprocal strategies but also for learned, meta, and finite-state policies.

6. Robustness, variants, and contemporary usage

A central lesson of recent work is that tournament outcomes are highly sensitive to field composition, noise, and termination assumptions. When the reproduced second-tournament field is expanded to 272 strategies—63 original submissions plus 209 additional Axelrod-Python entries—sophisticated learners trained via reinforcement learning dominate, including EvolvedLookerUp2_2_2, Evolved HMM 5, Evolved ANN, and PSO Gambler variants (Knight et al., 17 Oct 2025). Within that enlarged field, the strongest original strategy is k42r, ranked 16th with average score 2.735 and cooperation rate 0.820. This suggests that the original tournament was unusually favorable to TFT because the original field was highly cooperative.

The same point appears under targeted perturbation. Adding Stewart and Plotkin’s ZD strategies to the restored second-tournament field does not make extortion uniformly dominant; the top ranks remain highly cooperative. ZD-GTFT-2 with parameters $2R > T + S$9 leads with average score 2.834 and cooperation rate 0.926, followed by k42r at 2.825 and 0.903, GTFT: 0.33 at 2.819 and 0.941, and TFT at 2.812 and 0.888 (Knight et al., 17 Oct 2025). Under $T$0 noise, the mean cooperation rate of the original 63 strategies drops from 0.75 to 0.65, TFT no longer leads, and k42r ranks 10th while outperforming TFT. In the full Axelrod-Python set, several original entries rise to very high ranks under noise, with the best original ranking 5th at $T$1 and 3rd at $T$2 noise (Knight et al., 17 Oct 2025).

This robustness analysis also clarifies a common misconception. The classical conclusion was not that TFT is universally optimal, but that TFT was exceptionally strong in a highly cooperative and relatively clean environment. The restored second tournament shows that the field’s mean cooperation was approximately $T$3, with top strategies nearly perfectly cooperating with one another, and that many submitted strategies were designed to work well against TFT-like opponents (Knight et al., 17 Oct 2025). As the field becomes more diverse or noisy, strategies with extra mechanisms for escaping mutual punishment, guarding against exploitation, or adjusting to opponent cooperation rates do better. The behavior of k42r, described as an adaptive TFT-like design with modes that help recover from mutual defection and avoid alternating defections, is an explicit example (Knight et al., 17 Oct 2025).

For present-day use, the literature recommends explicit specification of payoffs, match lengths or termination hazards, noise, seeds, and field composition. If matching Axelrod’s second tournament, one may pre-sample five lengths such as $T$4 and apply them uniformly across pairings; if studying robustness, one should also evaluate noiseless and noisy variants, for example $T$5 and $T$6 action-flip noise (Knight et al., 17 Oct 2025). Reported metrics commonly include average score per strategy over all matches and realizations, average cooperation rates overall and pairwise, rank distributions, win proportions, and, in some frameworks, match-win counts (Knight et al., 2016). The broader reproducibility literature further recommends fixing random seeds, recording software versions and operating-system details, and using standardized open implementations (Knight et al., 2016).

The resulting methodology has become a general platform for comparative IPD research. It supports historical reconstruction, robustness analysis under diverse fields and noise regimes, and downstream ecological studies in which tournament payoffs feed Moran processes or replicator-style dynamics (Knight et al., 2016). In that sense, the Axelrod-style tournament is both a specific historical artifact and a durable experimental design: a reproducible population benchmark for studying how contingent cooperation, exploitation, adaptation, and noise interact in repeated social dilemmas.