Skip to main content
Browse Subject Areas

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Reinforcement learning produces dominant strategies for the Iterated Prisoner’s Dilemma

  • Marc Harper ,

    Contributed equally to this work with: Marc Harper, Vincent Knight

    Roles Conceptualization, Methodology, Software, Writing – original draft, Writing – review & editing

    Affiliation Google Inc., Mountain View, CA, United States of America

  • Vincent Knight ,

    Contributed equally to this work with: Marc Harper, Vincent Knight

    Roles Conceptualization, Data curation, Formal analysis, Methodology, Software, Visualization, Writing – original draft, Writing – review & editing

    Affiliation Cardiff University, School of Mathematics, Cardiff, United Kingdom

  • Martin Jones ,

    Roles Conceptualization, Software

    ‡These authors also contributed equally to this work.

    Affiliation Independent Researcher, Edinburgh, Scotland

  • Georgios Koutsovoulos ,

    Roles Conceptualization, Software

    ‡These authors also contributed equally to this work.

    Affiliation INRA, Université Côte d’Azur, CNRS, ISA, Nice, France

  • Nikoleta E. Glynatsi ,

    Roles Visualization, Writing – original draft

    ‡These authors also contributed equally to this work.

    Affiliation Cardiff University, School of Mathematics, Cardiff, United Kingdom

  • Owen Campbell

    Roles Software, Writing – review & editing

    ‡These authors also contributed equally to this work.

    Affiliation Independent Researcher, Chester, United Kingdom


We present tournament results and several powerful strategies for the Iterated Prisoner’s Dilemma created using reinforcement learning techniques (evolutionary and particle swarm algorithms). These strategies are trained to perform well against a corpus of over 170 distinct opponents, including many well-known and classic strategies. All the trained strategies win standard tournaments against the total collection of other opponents. The trained strategies and one particular human made designed strategy are the top performers in noisy tournaments also.


The Prisoner’s Dilemma (PD) is a two player game used to model a variety of strategic interactions. Each player chooses between cooperation (C) or defection (D). The payoffs of the game are defined by the matrix , where T > R > P > S and 2R > T + S. The PD is a one round game, but is commonly studied in a manner where the prior outcomes matter. This repeated form is called the Iterated Prisoner’s Dilemma (IPD). The IPD is frequently used to understand the evolution of cooperative behaviour from complex dynamics [1].

This manuscript uses the Axelrod library [2, 3], open source software for conducting IPD research with reproducibility as a principal goal. Written in the Python programming language, to date the library contains source code contributed by over 50 individuals from a variety of geographic locations and technical backgrounds. The library is supported by a comprehensive test suite that covers all the intended behaviors of all of the strategies in the library, as well as the features that conduct matches, tournaments, and population dynamics.

The library is continuously developed and as of version 3.0.0, the library contains over 200 strategies, many from the scientific literature, including classic strategies like Win Stay Lose Shift [4] and previous tournament winners such as OmegaTFT [5], Adaptive Pavlov [6], and ZDGTFT2 [7].

Since Robert Axelrod’s seminal tournament [8], a number of IPD tournaments have been undertaken and are summarised in Table 1. Further to the work described in [2] a regular set of standard, noisy [9] and probabilistic ending [10] tournaments are carried out as more strategies are added to the Axelrod library. Details and results are available here: This work presents a detailed analysis of tournaments with 176 strategies.

Table 1. An overview of a selection of published tournaments.

Not all tournaments were ‘standard’ round robins; for more details see the indicated references.

In this work we describe how collections of strategies in the Axelrod library have been used to train new strategies specifically to win IPD tournaments. These strategies are trained using generic strategy archetypes based on e.g. finite state machines, arriving at particularly effective parameter choices through evolutionary or particle swarm algorithms. There are several previous publications that use evolutionary algorithms to evolve IPD strategies in various circumstances [1322]. See also [23] for a strategy trained to win against a collection of well-known IPD opponents and see [24] for a prior use of particle swarm algorithms. Our results are unique in that we are able to train against a large and diverse collection of strategies available from the scientific literature. Crucially, the software used in this work is openly available and can be used to train strategies in the future in a reliable manner, with confidence that the opponent strategies are correctly implemented, tested and documented.

Materials and methods

The strategy archetypes

The Axelrod library now contains many parametrised strategies trained using machine learning methods. Most are deterministic, use many rounds of memory, and perform extremely well in tournaments as will be discussed in the results Section. Training will be discussed in a later section. These strategies can encode a variety of other strategies, including classic strategies like Tit For Tat [25], handshake strategies, and grudging strategies, that always defect after an opponent defection.


The LookerUp strategy is based on a lookup table and encodes a set of deterministic responses based on the opponent’s first n1 moves, the opponent’s last m1 moves, and the players last m2 moves. If n1 > 0 then the player has infinite memory depth, otherwise it has depth max(m1, m2). This is illustrated diagrammatically in Fig 1.

Fig 1. Diagrammatic representation of the looker up archetype.

Training of this strategy corresponds to finding maps from partial histories to actions, either a cooperation or a defection. Although various combinations of n1, m1, and m2 have been tried, the best performance at the time of training was obtained for n1 = m1 = m2 = 2 and generally for n1 > 0. A strategy called EvolvedLookerUp2_2_2 is among the top strategies in the library.

This archetype can be used to train deterministic memory-n strategies with the parameters n1 = 0 and m1 = m2 = n. For n = 1, the resulting strategy cooperates if the last round was mutual cooperation and defects otherwise, known as Grim or Grudger.

Two strategies in the library, Winner12 and Winner21, from [26], are based on lookup tables for n1 = 0, m1 = 1, and m2 = 2. The strategy Winner12 emerged in less than 10 generations of training in our framework using a score maximizing objective. Strategies nearly identical to Winner21 arise from training with a Moran process objective.


Gambler is a stochastic variant of LookerUp. Instead of deterministically encoded moves the lookup table emits probabilities which are used to choose cooperation or defection. This is illustrated diagrammatically in Fig 2.

Fig 2. Diagrammatic representation of the Gambler archetype.

Training of this strategy corresponds to finding maps from histories to a probability of cooperation. The library includes a strategy trained with n1 = m1 = m2 = 2 that is mostly deterministic, with 52 of the 64 probabilities being 0 or 1.

This strategy type can be used to train arbitrary memory-n strategies. A memory one strategy called PSOGamblerMem1 was trained, with probabilities (Pr(C | CC), Pr(C | CD), Pr(C | DC), Pr(C | DD)) = (1, 0.5217, 0, 0.121). Though it performs well in standard tournaments (see Table 2) it does not outperform the longer memory strategies, and is bested by a similar strategy that also uses the first round of play: PSOGambler_1_1_1.

Table 2. Standard tournament: Mean score per turn of top 15 strategies (ranked by median over 50000 tournaments).

The leaderboard is dominated by the trained strategies (indicated by a *).

These strategies are trained with a particle swarm algorithm rather than an evolutionary algorithm (though the former would suffice). Particle swarm algorithms have been used to trained IPD strategies previously [24].

ANN: Single hidden layer artificial neural network.

Strategies based on artificial neural networks use a variety of features computed from the history of play:

  • Opponent’s first move is C
  • Opponent’s first move is D
  • Opponent’s second move is C
  • Opponent’s second move is D
  • Player’s previous move is C
  • Player’s previous move is D
  • Player’s second previous move is C
  • Player’s second previous move is D
  • Opponent’s previous move is C
  • Opponent’s previous move is D
  • Opponent’s second previous move is C
  • Opponent’s second previous move is D
  • Total opponent cooperations
  • Total opponent defections
  • Total player cooperations
  • Total player defections
  • Round number

These are then input into a feed forward neural network with one layer and user-supplied width. This is illustrated diagrammatically in Fig 3.

Training of this strategy corresponds to finding parameters of the neural network. An inner layer with just five nodes performs quite well in both deterministic and noisy tournaments. The output of the ANN used in this work is deterministic; a stochastic variant that outputs probabilities rather than exact moves could be created.

Finite state machines.

Strategies based on finite state machines are deterministic and computationally efficient. In each round of play the strategy selects an action based on the current state and the opponent’s last action, transitioning to a new state for the next round. This is illustrated diagrammatically in Fig 4.

Fig 4. Diagrammatic representation of the finite state machine archetype.

Training this strategy corresponds to finding mappings of states and histories to an action and a state. Figs 5 and 6 show two of the trained finite state machines. The layout of state nodes is kept the same between Figs 5 and 6 to highlight the effect of different training environments. Note also that two of the 16 states are not used, this is also an outcome of the training process.

Fig 5. Evolved_FSM_16: Trained to maximize score in a standard tournament.

Fig 6. Evolved_FSM_16_Noise_05: Trained to maximize score in a noisy tournament.

Hidden markov models.

A variant of finite state machine strategies are called hidden Markov models (HMMs). Like the strategies based on finite state machines, these strategies also encode an internal state. However, they use probabilistic transitions based on the prior round of play to other states and cooperate or defect with various probabilities at each state. This is shown diagrammatically in Fig 7. Training this strategy corresponds to finding mappings of states and histories to probabilities of cooperating as well as probabilities of the next internal state.

Fig 7. Diagrammatic representation of the hidden markov model archetype.

Meta strategies.

There are several strategies based on ensemble methods that are common in machine learning called Meta strategies. These strategies are composed of a team of other strategies. In each round, each member of the team is polled for its desired next move. The ensemble then selects the next move based on a rule, such as the consensus vote in the case of MetaMajority or the best individual performance in the case of MetaWinner. These strategies were among the highest performing in the library before the inclusion of those trained by reinforcement learning.

Because these strategies inherit many of the properties of the strategies on which they are based, including using knowledge of the match length to defect on the last round(s) of play, not all of these strategies were included in results of this paper. These strategies do not typically outperform the trained strategies described above.

Training methods

The trained strategies (denoted by a * in Appendix A) were trained using reinforcement learning algorithms. The ideas of reinforcement learning can be attributed to the original work of [27] in which the notion that computers would learn by taking random actions but according to a distribution that picked actions with high rewards more often. The two particular algorithms used here:

  • Particle Swarm Algorithm: [28].
  • Evolutionary algorithm: [29].

The Particle Swarm Algorithm is implemented using the pyswarm library: This algorithm was used only to train the Gambler archetype.

All other strategies were trained using evolutionary algorithms. The evolutionary algorithms used standard techniques, varying strategies by mutation and crossover, and evaluating the performance against each opponent for many repetitions. The best performing strategies in each generation are persisted, variants created, and objective functions computed again.

The default parameters for this procedure:

  • A population size of 40 individuals (kept constant across the generations);
  • A mutation rate of 10%;
  • 10 individuals kept from one generation to the next;
  • A total of 500 generations.

All implementations of these algorithms are archived at [30]. This software is (similarly to the Axelrod library) available on github There are objective functions for:

  • total or mean payoff,
  • total or mean payoff difference (unused in this work),
  • total Moran process wins (fixation probability). This lead to the strategies named TF1, TF2, TF3 listed in Appendix A.

These can be used in noisy or standard environments. These objectives can be further modified to suit other purposes. New strategies could be trained with variations including spatial structure and probabilistically ending matches.


This section presents the results of a large IPD tournament with strategies from the Axelrod library, including some additional parametrized strategies (e.g. various parameter choices for Generous Tit For Tat [23]). These are listed in Appendix A.

All strategies in the tournament follow a simple set of rules in accordance with earlier tournaments:

  • Players are unaware of the number of turns in a match.
  • Players carry no acquired state between matches.
  • Players cannot observe the outcome of other matches.
  • Players cannot identify their opponent by any label or identifier.
  • Players cannot manipulate or inspect their opponents in any way.

Any strategy that does not follow these rules, such as a strategy that defects on the last round of play, was omitted from the tournament presented here (but not necessarily from the training pool).

A total of 176 are included, of which 53 are stochastic. A standard tournament with 200 turns and a tournament with 5% noise is discussed. Due to the inherent stochasticity of these IPD tournaments, these tournaments were repeated 50000 times. This allows for a detailed and confident analysis of the performance of strategies. To illustrate the results considered, Fig 8 shows the distribution of the mean score per turn of Tit For Tat over all the repetitions. Similarly, Fig 9 shows the ranks of of Tit For Tat for each repetition. (We note that it never wins a tournament). Finally Fig 10 shows the number of opponents beaten in any given tournament: Tit For Tat does not win any match. (This is due to the fact that it will either draw with mutual cooperation or defect second).

The utilities used are (R, P, T, S) = (3, 1, 5, 0) thus the specific Prisoner’s Dilemma being played is: (1)

All data generated for this work is archived and available at [31].

Standard tournament

The top 11 performing strategies by median payoff are all strategies trained to maximize total payoff against a subset of the strategies (Table 2). The next strategy is Desired Belief Strategy (DBS) [32], which actively analyzes the opponent and responds accordingly. The next two strategies are Winner12, based on a lookup table, Fool Me Once [3], a grudging strategy that defects indefinitely on the second defection, and Omega Tit For Tat [12].

For completeness, violin plots showing the distribution of the scores of each strategy (again ranked by median score) are shown in Fig 11.

Fig 11. Standard tournament: Mean score per turn (strategies ordered by median score over 50000 tournaments).

Pairwise payoff results are given as a heatmap (Fig 12) which shows that many strategies achieve mutual cooperation (obtaining a score of 3). The top performing strategies never defect first yet are able to exploit weaker strategies that attempt to defect.

Fig 12. Standard tournament: Mean score per turn of row players against column players (ranked by median over 50000 tournaments).

The strategies that win the most matches (Table 3) are Defector [1] and Aggravater [3], followed by handshaking and zero determinant strategies [33]. This includes two handshaking strategies that were the result of training to maximize Moran process fixation (TF1 and TF2). No strategies were trained specifically to win matches. None of the top scoring strategies appear in the top 15 list of strategies ranked by match wins. This can be seen in Fig 13 where the distribution of the number of wins of each strategy is shown.

Table 3. Standard tournament: Number of wins per tournament of top 15 strategies (ranked by median wins over 50000 tournaments).

Fig 13. Standard tournament: Number of wins per tournament (ranked by median over 50000 tournaments).

The number of wins of the top strategies of Table 2 are shown in Table 4. It is evident that although these strategies score highly they do not win many matches: the strategy with the most number of wins is the Evolved FSM 16 strategy that at most won 60 (60/175 ≈ 34%) matches in a given tournament.

Table 4. Standard tournament: Number of wins per tournament of top 15 strategies (ranked by median score over 50000 tournaments) * indicates that the strategy was trained.

Finally, Table 5 and Fig 14 show the ranks (based on median score) of each strategy over the repeated tournaments. Whilst there is some stochasticity, the top three strategies almost always rank in the top three. For example, the worst that the EvolvedLookerUp_2_2_2 ranks in any tournament is 8th.

Table 5. Standard tournament: Rank in each tournament of top 15 strategies (ranked by median over 50000 tournaments) * indicates that the strategy was trained.

Fig 14. Standard tournament: Rank in each tournament (ranked by median over 50000 tournaments).

Figs 1517 shows the rate of cooperation in each round for the top three strategies. The opponents in these figures are ordered according to performance by median score. It is evident that the high performing strategies share a common thread against the top strategies: they do not defect first and achieve mutual cooperation. Against the lower strategies they also do not defect first (a mean cooperation rate of 1 in the first round) but do learn to quickly retaliate.

Fig 15. Cooperation rates for EvolvedLookerUp_2_2_2 (strategies ordered by median score over 10000 tournaments).

Fig 16. Cooperation rates for Evolved_HMM_5 (strategies ordered by median score over 10000 tournaments).

Fig 17. Cooperation rates for Evolved_FSM_16 (strategies ordered by median score over 10000 tournaments).

Noisy tournament

Results from noisy tournaments in which there is a 5% chance that an action is flipped are now described. As shown in Table 6 and Fig 18, the best performing strategies in median payoff are DBS, designed to account for noise, followed by two strategies trained in the presence of noise and three trained strategies trained without noise. One of the strategies trained with noise (PSO Gambler) actually performs less well than some of the other high ranking strategies including Spiteful TFT (TFT but defects indefinitely if the opponent defects twice consecutively) and OmegaTFT (also designed to handle noise). While DBS is the clear winner, it comes at a 6x increased run time over Evolved FSM 16 Noise 05.

Table 6. Noisy (5%) tournament: Mean score per turn of top 15 strategies (ranked by median over 50000 tournaments) * indicates that the strategy was trained.

Fig 18. Noisy (5%) tournament: Mean score per turn (strategies ordered by median score over 50000 tournaments).

Recalling Table 2, the strategies trained in the presence of noise are also among the best performers in the absence of noise. As shown in Fig 19 the cluster of mutually cooperative strategies is broken by the noise at 5%. A similar collection of players excels at winning matches but again they have a poor total payoff.

Fig 19. Noisy (5%) tournament: Mean score per turn of row players against column players (ranked by median over 50000 tournaments).

As shown in Table 7 and Fig 20 the strategies tallying the most wins are somewhat similar to the standard tournaments, with Defector, the handshaking CollectiveStrategy [34], and Aggravater appearing as the top three again.

Table 7. Noisy (5%) tournament: Number of wins per tournament of top 15 strategies (ranked by median wins over 50000 tournaments).

Fig 20. Noisy (5%) tournament: Number of wins per tournament (strategies ordered by median score over 50000 tournaments).

As shown in Table 8, the top ranking strategies win a larger number of matches in the presence of noise. For example Spiteful Tit For Tat [35] in one tournament won almost all its matches (167).

Table 8. Noisy (5%) tournament: Number of wins per tournament of top 15 strategies (ranked by median score over 50000 tournaments) * indicates that the strategy was trained.

Finally, Table 9 and Fig 21 show the ranks (based on median score) of each strategy over the repeated tournaments. We see that the stochasticity of the ranks understandably increases relative to the standard tournament. An exception is the top three strategies, for example, the DBS strategy never ranks lower than second and wins 75% of the time. The two strategies trained for noisy tournaments rank in the top three 95% of the time.

Table 9. Noisy (5%) tournament: Rank in each tournament of top 15 strategies (ranked by median over 50000 tournaments) * indicates that the strategy was trained.

Fig 21. Noisy (5%) tournament: Rank in each tournament (strategies ordered by median score over 50000 tournaments).

Figs 2224 shows the rate of cooperation in each round for the top three strategies (in the absence of noise) and just as for the top performing strategies in the standard tournament it is evident that the strategies never defect first and learn to quickly punish poorer strategies.

Fig 22. Cooperation rates for DBS (strategies ordered by median score over 10000 tournaments).

Fig 23. Cooperation rates for Evolved_ANN_5_Noise_05 (strategies ordered by median score over 10000 tournaments).

Fig 24. Cooperation rates for Evolved_FSM_16_Noise_05 (strategies ordered by median score over 10000 tournaments).


The tournament results indicate that pre-trained strategies are generally better than human designed strategies at maximizing payoff against a diverse set of opponents. An evolutionary algorithm produces strategies based on multiple generic archetypes that are able to achieve a higher average score than any other known opponent in a standard tournament. Most of the trained strategies use multiple rounds of the history of play (some using all of it) and outperform memory-one strategies from the literature. Interestingly, a trained memory one strategy produced by a particle swarm algorithm performs well, better than human designed strategies such as Win Stay Lose Shift and zero determinant strategies (which enforce a payoff difference rather than maximize total payoff).

In opposition to historical tournament results and community folklore, our results show that complex strategies can be effective for the IPD. Of all the human-designed strategies in the library, only DBS consistently performs well, and it is substantially more complex than traditional tournament winners like TFT, OmegaTFT, and zero determinant strategies.

The generic structure of the trained strategies did not appear to be critical for the standard tournament—strategies based on lookup tables, finite state machines, neural networks, and stochastic variants all performed well. Single layer neural networks performed well in both noisy and standard tournaments though these had some aspect of human involvement in the selection of features. This is in line with the other strategies also where some human decisions are made regarding the structure. For the LookerUp and Gambler archetypes a decision has to be made regarding the number of rounds of history and initial play that are to be used. In contrast, the finite state machines and hidden Markov models required only a choice of the number of states, and the training algorithm can eliminate unneeded states in the case of finite state machines (evidenced by the unconnected nodes in the diagrams for the included representations).

Many strategies can be represented by multiple archetypes, however some archetypes will be more efficient in encoding the patterns present in the data. The fact that the Lookerup strategy does the best for the standard tournament indicates that it represents an efficient reduction of dimension which in turn makes its training more efficient. In particular the first rounds of play were valuable bits of information. For the noisy tournament however the dimension reduction represented by some archetypes indicates that some features of the data are not captured by the lookup tables while they are by the neural networks and the finite state machines, allowing the latter to adapt better to the noisy environment. Intuitively, a noisy environment can significantly affect a lookup table based on the last two rounds of play since these action pairs compete with probing defections, apologies, and retaliations. Accordingly, it is not surprising that additional parameter space is needed to adapt to a noisy environment.

Two strategies designed specifically to account for noise, DBS and OmegaTFT, perform well and only DBS performs better than the trained strategies and only in noisy contexts. Empirically we find that DBS (with its default parameters) does not win tournaments at 1% noise. However DBS has a parameter that accounts for the expected amount of noise and a followup study with various noise levels could make a more complete study of the performance of DBS and strategies trained at various noise levels.

The strategies trained to maximize their average score are generally cooperative and do not defect first. Maximizing for individual performance across a collection of opponents leads to mutual cooperation despite the fact that mutual cooperation is an unstable evolutionary equilibrium for the prisoner’s dilemma. Specifically it is noted that the reinforcement learning process for maximizing payoff does not lead to exploitative zero determinant strategies, which may also be a result of the collection of training strategies, of which several retaliate harshly. Training with the objective of maximizing payoff difference may produce strategies more like zero determinant strategies.

For the trained strategies utilizing look up tables we generally found those that incorporate one or more of the initial rounds of play outperformed those that did not. The strategies based on neural networks and finite state machines also are able to condition throughout a match on the first rounds of play. Accordingly, we conclude that first impressions matter in the IPD. The best strategies are nice (never defecting first) and the impact of the first rounds of play could be further investigated with the Axelrod library in future work by e.g. forcing all strategies to defect on the first round.

We note that as the library grows, the top performing strategies sometimes shuffle, and are not retrained automatically. Most of the strategies were trained on an earlier version of the library (v2.2.0: [36]) that did not include DBS and several other opponents. The precise parameters that are optimal will depend on the pool of opponents. Moreover we have not extensively trained strategies to determine the minimum parameter spaces that are sufficient—neural networks with fewer nodes and features and finite state machines with fewer states may suffice. See [37] for discussion of resource availability for IPD strategies.

Finally, whilst we have considered the robustness of our claims and results with respect to noise it would also be of interest to train strategies for different versions of the stage game (also referred to as dilemma strength) [38, 39]. Our findings seems to indicate that obtaining strong strategies for other games through reinforcement learning would be possible.

A Appendix A: List of players

The players used for this study are from Axelrod version 2.13.0 [3].

  1. ϕDeterministicMemory depth: ∞. [3]
  2. πDeterministicMemory depth: ∞. [3]
  3. eDeterministicMemory depth: ∞. [3]
  4. ALLCorALLD—StochasticMemory depth: 1. [3]
  5. Adaptive—DeterministicMemory depth: ∞. [43]
  6. Adaptive Pavlov 2006—DeterministicMemory depth: ∞. [12]
  7. Adaptive Pavlov 2011—DeterministicMemory depth: ∞. [43]
  8. Adaptive Tit For Tat: 0.5—DeterministicMemory depth: ∞. [44]
  9. Aggravater—DeterministicMemory depth: ∞. [3]
  10. Alternator—DeterministicMemory depth: 1. [1, 45]
  11. Alternator Hunter—DeterministicMemory depth: ∞. [3]
  12. Anti Tit For Tat—DeterministicMemory depth: 1. [46]
  13. AntiCycler—DeterministicMemory depth: ∞. [3]
  14. Appeaser—DeterministicMemory depth: ∞. [3]
  15. Arrogant QLearner—StochasticMemory depth: ∞. [3]
  16. Average Copier—StochasticMemory depth: ∞. [3]
  17. Better and Better—StochasticMemory depth: ∞. [35]
  18. Bully—DeterministicMemory depth: 1. [47]
  19. Calculator—StochasticMemory depth: ∞. [35]
  20. Cautious QLearner—StochasticMemory depth: ∞. [3]
  21. CollectiveStrategy (CS)—DeterministicMemory depth: ∞. [34]
  22. Contrite Tit For Tat (CTfT)—DeterministicMemory depth: 3. [48]
  23. Cooperator—DeterministicMemory depth: 0. [1, 33, 45]
  24. Cooperator Hunter—DeterministicMemory depth: ∞. [3]
  25. Cycle Hunter—DeterministicMemory depth: ∞. [3]
  26. Cycler CCCCCD—DeterministicMemory depth: 5. [3]
  27. Cycler CCCD—DeterministicMemory depth: 3. [3]
  28. Cycler CCCDCD—DeterministicMemory depth: 5. [3]
  29. Cycler CCD—DeterministicMemory depth: 2. [45]
  30. Cycler DC—DeterministicMemory depth: 1. [3]
  31. Cycler DDC—DeterministicMemory depth: 2. [45]
  32. DBS: 0.75, 3, 4, 3, 5—DeterministicMemory depth: ∞. [32]
  33. Davis: 10—DeterministicMemory depth: ∞. [25]
  34. Defector—DeterministicMemory depth: 0. [1, 33, 45]
  35. Defector Hunter—DeterministicMemory depth: ∞. [3]
  36. Desperate—StochasticMemory depth: 1. [49]
  37. DoubleResurrection—DeterministicMemory depth: 5. [50]
  38. Doubler—DeterministicMemory depth: ∞. [35]
  39. Dynamic Two Tits For Tat—StochasticMemory depth: 2. [3]
  40. EasyGo—DeterministicMemory depth: ∞. [35, 43]
  41. Eatherley—StochasticMemory depth: ∞. [10]
  42. Eventual Cycle Hunter—DeterministicMemory depth: ∞. [3]
  43. Evolved ANN—DeterministicMemory depth: ∞. [3]
  44. Evolved ANN 5—DeterministicMemory depth: ∞. [3]
  45. Evolved ANN 5 Noise 05—DeterministicMemory depth: ∞. [3]
  46. Evolved FSM 16—DeterministicMemory depth: 16. [3]
  47. Evolved FSM 16 Noise 05—DeterministicMemory depth: 16. [3]
  48. Evolved FSM 4—DeterministicMemory depth: 4. [3]
  49. Evolved HMM 5—StochasticMemory depth: 5. [3]
  50. EvolvedLookerUp1_1_1—DeterministicMemory depth: ∞. [3]
  51. EvolvedLookerUp2_2_2—DeterministicMemory depth: ∞. [3]
  52. Feld: 1.0, 0.5, 200—StochasticMemory depth: 200. [25]
  53. Firm But Fair—StochasticMemory depth: 1. [51]
  54. Fool Me Forever—DeterministicMemory depth: ∞. [3]
  55. Fool Me Once—DeterministicMemory depth: ∞. [3]
  56. Forgetful Fool Me Once: 0.05—StochasticMemory depth: ∞. [3]
  57. Forgetful Grudger—DeterministicMemory depth: 10. [3]
  58. Forgiver—DeterministicMemory depth: ∞. [3]
  59. Forgiving Tit For Tat (FTfT)—DeterministicMemory depth: ∞. [3]
  60. Fortress3—DeterministicMemory depth: 3. [14]
  61. Fortress4—DeterministicMemory depth: 4. [14]
  62. GTFT: 0.1—StochasticMemory depth: 1.
  63. GTFT: 0.3—StochasticMemory depth: 1.
  64. GTFT: 0.33—StochasticMemory depth: 1. [23, 52]
  65. GTFT: 0.7—StochasticMemory depth: 1.
  66. GTFT: 0.9—StochasticMemory depth: 1.
  67. General Soft Grudger: n = 1, d = 4, c = 2—DeterministicMemory depth: ∞. [3]
  68. Gradual—DeterministicMemory depth: ∞. [53]
  69. Gradual Killer: (‘D’, ‘D’, ‘D’, ‘D’, ‘D’, ‘C’, ‘C’)—DeterministicMemory depth: ∞. [35]
  70. Grofman—StochasticMemory depth: ∞. [25]
  71. Grudger—DeterministicMemory depth: 1. [25, 43, 49, 53, 54]
  72. GrudgerAlternator—DeterministicMemory depth: ∞. [35]
  73. Grumpy: Nice, 10, −10—DeterministicMemory depth: ∞. [3]
  74. Handshake—DeterministicMemory depth: ∞. [55]
  75. Hard Go By Majority—DeterministicMemory depth: ∞. [45]
  76. Hard Go By Majority: 10—DeterministicMemory depth: 10. [3]
  77. Hard Go By Majority: 20—DeterministicMemory depth: 20. [3]
  78. Hard Go By Majority: 40—DeterministicMemory depth: 40. [3]
  79. Hard Go By Majority: 5—DeterministicMemory depth: 5. [3]
  80. Hard Prober—DeterministicMemory depth: ∞. [35]
  81. Hard Tit For 2 Tats (HTf2T)—DeterministicMemory depth: 3. [7]
  82. Hard Tit For Tat (HTfT)—DeterministicMemory depth: 3. [56]
  83. Hesitant QLearner—StochasticMemory depth: ∞. [3]
  84. Hopeless—StochasticMemory depth: 1. [49]
  85. Inverse—StochasticMemory depth: ∞. [3]
  86. Inverse Punisher—DeterministicMemory depth: ∞. [3]
  87. Joss: 0.9—StochasticMemory depth: 1. [7, 25]
  88. Level Punisher—DeterministicMemory depth: ∞. [50]
  89. Limited Retaliate 2: 0.08, 15—DeterministicMemory depth: ∞. [3]
  90. Limited Retaliate 3: 0.05, 20—DeterministicMemory depth: ∞. [3]
  91. Limited Retaliate: 0.1, 20—DeterministicMemory depth: ∞. [3]
  92. MEM2—DeterministicMemory depth: ∞. [57]
  93. Math Constant Hunter—DeterministicMemory depth: ∞. [3]
  94. Meta Hunter Aggressive: 7 players—DeterministicMemory depth: ∞. [3]
  95. Meta Hunter: 6 players—DeterministicMemory depth: ∞. [3]
  96. Meta Mixer: 173 players—StochasticMemory depth: ∞. [3]
  97. Naive Prober: 0.1—StochasticMemory depth: 1. [43]
  98. Negation—StochasticMemory depth: 1. [56]
  99. Nice Average Copier—StochasticMemory depth: ∞. [3]
  100. Nydegger—DeterministicMemory depth: 3. [25]
  101. Omega TFT: 3, 8—DeterministicMemory depth: ∞. [12]
  102. Once Bitten—DeterministicMemory depth: 12. [3]
  103. Opposite Grudger—DeterministicMemory depth: ∞. [3]
  104. PSO Gambler 1_1_1—StochasticMemory depth: ∞. [3]
  105. PSO Gambler 2_2_2—StochasticMemory depth: ∞. [3]
  106. PSO Gambler 2_2_2 Noise 05—StochasticMemory depth: ∞. [3]
  107. PSO Gambler Mem1—StochasticMemory depth: 1. [3]
  108. Predator—DeterministicMemory depth: 9. [14]
  109. Prober—DeterministicMemory depth: ∞. [43]
  110. Prober 2—DeterministicMemory depth: ∞. [35]
  111. Prober 3—DeterministicMemory depth: ∞. [35]
  112. Prober 4—DeterministicMemory depth: ∞. [35]
  113. Pun1—DeterministicMemory depth: 2. [14]
  114. Punisher—DeterministicMemory depth: ∞. [3]
  115. Raider—DeterministicMemory depth: 3. [17]
  116. Random Hunter—DeterministicMemory depth: ∞. [3]
  117. Random: 0.1—StochasticMemory depth: 0.
  118. Random: 0.3—StochasticMemory depth: 0.
  119. Random: 0.5—StochasticMemory depth: 0. [25, 44]
  120. Random: 0.7—StochasticMemory depth: 0.
  121. Random: 0.9—StochasticMemory depth: 0.
  122. Remorseful Prober: 0.1—StochasticMemory depth: 2. [43]
  123. Resurrection—DeterministicMemory depth: 5. [50]
  124. Retaliate 2: 0.08—DeterministicMemory depth: ∞. [3]
  125. Retaliate 3: 0.05—DeterministicMemory depth: ∞. [3]
  126. Retaliate: 0.1—DeterministicMemory depth: ∞. [3]
  127. Revised Downing: True—DeterministicMemory depth: ∞. [25]
  128. Ripoff—DeterministicMemory depth: 2. [58]
  129. Risky QLearner—StochasticMemory depth: ∞. [3]
  130. SelfSteem—StochasticMemory depth: ∞. [59]
  131. ShortMem—DeterministicMemory depth: 10. [59]
  132. Shubik—DeterministicMemory depth: ∞. [25]
  133. Slow Tit For Two Tats—DeterministicMemory depth: 2. [3]
  134. Slow Tit For Two Tats 2—DeterministicMemory depth: 2. [35]
  135. Sneaky Tit For Tat—DeterministicMemory depth: ∞. [3]
  136. Soft Go By Majority—DeterministicMemory depth: ∞. [1, 45]
  137. Soft Go By Majority: 10—DeterministicMemory depth: 10. [3]
  138. Soft Go By Majority: 20—DeterministicMemory depth: 20. [3]
  139. Soft Go By Majority: 40—DeterministicMemory depth: 40. [3]
  140. Soft Go By Majority: 5—DeterministicMemory depth: 5. [3]
  141. Soft Grudger—DeterministicMemory depth: 6. [43]
  142. Soft Joss: 0.9—StochasticMemory depth: 1. [35]
  143. SolutionB1—DeterministicMemory depth: 3. [15]
  144. SolutionB5—DeterministicMemory depth: 5. [15]
  145. Spiteful Tit For Tat—DeterministicMemory depth: ∞. [35]
  146. Stochastic Cooperator—StochasticMemory depth: 1. [60]
  147. Stochastic WSLS: 0.05—StochasticMemory depth: 1. [3]
  148. Suspicious Tit For Tat—DeterministicMemory depth: 1. [46, 53]
  149. TF1—DeterministicMemory depth: ∞. [3]
  150. TF2—DeterministicMemory depth: ∞. [3]
  151. TF3—DeterministicMemory depth: ∞. [3]
  152. Tester—DeterministicMemory depth: ∞. [10]
  153. ThueMorse—DeterministicMemory depth: ∞. [3]
  154. ThueMorseInverse—DeterministicMemory depth: ∞. [3]
  155. Thumper—DeterministicMemory depth: 2. [58]
  156. Tit For 2 Tats (Tf2T)—DeterministicMemory depth: 2. [1]
  157. Tit For Tat (TfT)—DeterministicMemory depth: 1. [25]
  158. Tricky Cooperator—DeterministicMemory depth: 10. [3]
  159. Tricky Defector—DeterministicMemory depth: ∞. [3]
  160. Tullock: 11—StochasticMemory depth: 11. [25]
  161. Two Tits For Tat (2TfT)—DeterministicMemory depth: 2. [1]
  162. VeryBad—DeterministicMemory depth: ∞. [59]
  163. Willing—StochasticMemory depth: 1. [49]
  164. Win-Shift Lose-Stay: D (WShLSt)—DeterministicMemory depth: 1. [43]
  165. Win-Stay Lose-Shift: C (WSLS)—DeterministicMemory depth: 1. [7, 52, 61]
  166. Winner12—DeterministicMemory depth: 2. [26]
  167. Winner21—DeterministicMemory depth: 2. [26]
  168. Worse and Worse—StochasticMemory depth: ∞. [35]
  169. Worse and Worse 2—StochasticMemory depth: ∞. [35]
  170. Worse and Worse 3—StochasticMemory depth: ∞. [35]
  171. ZD-Extort-2 v2: 0.125, 0.5, 1—StochasticMemory depth: 1. [62]
  172. ZD-Extort-2: 0.1111111111111111, 0.5—StochasticMemory depth: 1. [7]
  173. ZD-Extort-4: 0.23529411764705882, 0.25, 1—StochasticMemory depth: 1. [3]
  174. ZD-GEN-2: 0.125, 0.5, 3—StochasticMemory depth: 1. [62]
  175. ZD-GTFT-2: 0.25, 0.5—StochasticMemory depth: 1. [7]
  176. ZD-SET-2: 0.25, 0.0, 2—StochasticMemory depth: 1. [62]


This work was performed using the computational facilities of the Advanced Research Computing @ Cardiff (ARCCA) Division, Cardiff University.

A variety of software libraries have been used in this work:

  • The Axelrod library (IPD strategies and Tournaments) [3].
  • The matplotlib library (visualisation) [40].
  • The pandas and numpy libraries (data manipulation) [41, 42].


  1. 1. Axelrod RM. The evolution of cooperation. Basic books; 2006.
  2. 2. Knight V, Campbell O, Harper M, Langner K, Campbell J, Campbell T, et al. An Open Framework for the Reproducible Study of the Iterated Prisoner’s Dilemma. Journal of Open Research Software. 2016;4(1).
  3. 3. project developers TA. Axelrod-Python/Axelrod: v2.13.0; 2017.
  4. 4. Nowak M, Sigmund K. A strategy of win-stay, lose-shift that outperforms tit-for-tat in the Prisoner’s Dilemma game. Nature. 1993;364(6432):56. pmid:8316296
  5. 5. Slany W, Kienreich W. On some winning strategies for the Iterated Prisoner’s Dilemma, or, Mr. Nice Guy and the Cosa Nostra. The Iterated Prisoners’ Dilemma: 20 Years on. 2007;4:171.
  6. 6. Li J. How to design a strategy to win an IPD tournament. The iterated prisoner’s dilemma. 2007;20:89–104.
  7. 7. Stewart AJ, Plotkin JB. Extortion and cooperation in the Prisoner’s Dilemma. Proceedings of the National Academy of Sciences. 2012;109(26):10134–10135.
  8. 8. Axelrod R. Effective Choice in the Prisoner’s Dilemma. Journal of Conflict Resolution. 1980;24(1):3–25.
  9. 9. Bendor J, Kramer RM, Stout S. When in doubt…: Cooperation in a noisy prisoner’s dilemma. Journal of Conflict Resolution. 1991;35(4):691–719.
  10. 10. Axelrod R. More Effective Choice in the Prisoner’s Dilemma. Journal of Conflict Resolution. 1980;24(3):379–403.
  11. 11. Stephens DW, McLinn CM, Stevens JR. Discounting and reciprocity in an Iterated Prisoner’s Dilemma. Science (New York, NY). 2002;298(5601):2216–2218.
  12. 12. Kendall G, Yao X, Chong SY. The iterated prisoners’ dilemma: 20 years on. vol. 4. World Scientific; 2007.
  13. 13. Ashlock D. Training function stacks to play the iterated prisoner’s dilemma. In: Computational Intelligence and Games, 2006 IEEE Symposium on. IEEE; 2006. p. 111–118.
  14. 14. Ashlock W, Ashlock D. Changes in prisoner’s dilemma strategies over evolutionary time with different population sizes. In: Evolutionary Computation, 2006. CEC 2006. IEEE Congress on. IEEE; 2006. p. 297–304.
  15. 15. Ashlock D, Brown JA, Hingston P. Multiple Opponent Optimization of Prisoner’s Dilemma Playing Agents. IEEE Transactions on Computational Intelligence and AI in Games. 2015;7(1):53–65.
  16. 16. Ashlock W, Ashlock D. Shaped prisoner’s dilemma automata. In: Computational Intelligence and Games (CIG), 2014 IEEE Conference on. IEEE; 2014. p. 1–8.
  17. 17. Ashlock W, Tsang J, Ashlock D. The evolution of exploitation. In: Foundations of Computational Intelligence (FOCI), 2014 IEEE Symposium on. IEEE; 2014. p. 135–142.
  18. 18. Barlow LA, Ashlock D. Varying decision inputs in Prisoner’s Dilemma. In: Computational Intelligence in Bioinformatics and Computational Biology (CIBCB), 2015 IEEE Conference on. IEEE; 2015. p. 1–8.
  19. 19. Fogel DB. Evolving behaviors in the iterated prisoner’s dilemma. Evolutionary Computation. 1993;1(1):77–97.
  20. 20. Marks RE. Niche strategies: the Prisoner’s Dilemma computer tournaments revisited. In: JOURNAL OF EVOLUTIONARY ECONOMICS. Citeseer; 1989.
  21. 21. Sudo T, Goto K, Nojima Y, Ishibuchi H. Effects of ensemble action selection with different usage of player’s memory resource on the evolution of cooperative strategies for iterated prisoner’s dilemma game. In: Evolutionary Computation (CEC), 2015 IEEE Congress on. IEEE; 2015. p. 1505–1512.
  22. 22. Vassiliades V, Christodoulou C. Multiagent reinforcement learning in the iterated prisoner’s dilemma: fast cooperation through evolved payoffs. In: Neural Networks (IJCNN), The 2010 International Joint Conference on. IEEE; 2010. p. 1–8.
  23. 23. Gaudesi M, Piccolo E, Squillero G, Tonda A. Exploiting evolutionary modeling to prevail in iterated prisoner’s dilemma tournaments. IEEE Transactions on Computational Intelligence and AI in Games. 2016;8(3):288–300.
  24. 24. Franken N, Engelbrecht AP. Particle swarm optimization approaches to coevolve strategies for the iterated prisoner’s dilemma. IEEE Transactions on Evolutionary Computation. 2005;9(6):562–579.
  25. 25. Axelrod R. Effective choice in the prisoner’s dilemma. Journal of conflict resolution. 1980;24(1):3–25.
  26. 26. Mathieu P, Delahaye JP. New Winning Strategies for the Iterated Prisoner’s Dilemma (Extended Abstract). 14th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2015). 2015; p. 1665–1666.
  27. 27. Turing AM. Computing machinery and intelligence. Mind. 1950;59(236):433–460.
  28. 28. Imran M, Hashim R, Khalid NEA. An overview of particle swarm optimization variants. Procedia Engineering. 2013;53:491–496.
  29. 29. Moriarty DE, Schultz AC, Grefenstette JJ. Evolutionary algorithms for reinforcement learning. J Artif Intell Res(JAIR). 1999;11:241–276.
  30. 30. Harper M, Knight V, Jones M, Koutsovoulos G. Axelrod-Python/axelrod-dojo: V0.0.2; 2017.
  31. 31. Knight V, Harper M. Data for: Reinforcement Learning Produces Dominant Strategies for the Iterated Prisoner’s Dilemma; 2017.
  32. 32. Au TC, Nau D. Accident or intention: that is the question (in the Noisy Iterated Prisoner’s Dilemma). In: Proceedings of the fifth international joint conference on Autonomous agents and multiagent systems. ACM; 2006. p. 561–568.
  33. 33. Press WH, Dyson FJ. Iterated Prisoner’s Dilemma contains strategies that dominate any evolutionary opponent. Proceedings of the National Academy of Sciences of the United States of America. 2012;109(26):10409–13. pmid:22615375
  34. 34. Li J, Kendall G. A strategy with novel evolutionary features for the iterated prisoner’s dilemma. Evolutionary Computation. 2009;17(2):257–274. pmid:19413490
  35. 35. LIFL. PRISON; 2008.
  36. 36. project developers TA. Axelrod-Python/Axelrod: v2.2.0; 2016.
  37. 37. Ashlock D, Kim EY. The impact of varying resources available to iterated prisoner’s dilemma agents. In: Foundations of Computational Intelligence (FOCI), 2013 IEEE Symposium on. IEEE; 2013. p. 60–67.
  38. 38. Wang Z, Kokubo S, Jusup M, Tanimoto Jun. Universal scaling for the dilemma strength in evolutionary games. Physics of life reviews. 2015;(14):1–30.
  39. 39. Tanimoto J, Sagara H. Relationship between dilemma occurrence and the existence of a weakly dominant strategy in a two-player symmetric game. BioSystems. 2007 Aug 31;90(1):105–14. pmid:17188808
  40. 40. Hunter JD. Matplotlib: A 2D graphics environment. Computing In Science & Engineering. 2007;9(3):90–95.
  41. 41. McKinney W, et al. Data structures for statistical computing in python. In: Proceedings of the 9th Python in Science Conference. vol. 445. van der Voort S, Millman J; 2010. p. 51–56.
  42. 42. Walt Svd, Colbert SC, Varoquaux G. The NumPy array: a structure for efficient numerical computation. Computing in Science & Engineering. 2011;13(2):22–30.
  43. 43. Li J, Hingston P, Member S, Kendall G. Engineering Design of Strategies for Winning Iterated Prisoner’ s Dilemma Competitions. 2011;3(4):348–360.
  44. 44. Tzafestas E. Toward adaptive cooperative behavior. From Animals to animals: Proceedings of the 6th International Conference on the Simulation of Adaptive Behavior (SAB-2000). 2000;2:334–340.
  45. 45. Mittal S, Deb K. Optimal strategies of the iterated prisoner’s dilemma problem for multiple conflicting objectives. IEEE Transactions on Evolutionary Computation. 2009;13(3):554–565.
  46. 46. Hilbe C, Nowak MA, Traulsen A. Adaptive dynamics of extortion and compliance. PloS one. 2013;8(11):e77886. pmid:24223739
  47. 47. Nachbar JH. Evolution in the finitely repeated prisoner’s dilemma. Journal of Economic Behavior & Organization. 1992;19(3):307–326.
  48. 48. Wu J, Axelrod R. How to cope with noise in the iterated prisoner’s dilemma. Journal of Conflict resolution. 1995;39(1):183–189.
  49. 49. van den Berg P, Weissing FJ. The importance of mechanisms for the evolution of cooperation. In: Proc. R. Soc. B. vol. 282. The Royal Society; 2015. p. 20151382.
  50. 50. Arnold E. CoopSim v0.9.9 beta 6; 2015.
  51. 51. Frean MR. The prisoner’s dilemma without synchrony. Proceedings of the Royal Society of London B: Biological Sciences. 1994;257(1348):75–79.
  52. 52. Nowak M, Sigmund K. A strategy of win-stay, lose-shift that outperforms tit-for-tat in the Prisoner’s Dilemma game. Nature. 1993;364(6432):56–58. pmid:8316296
  53. 53. Beaufils B, Delahaye JP, Mathieu P. Our meeting with gradual, a good strategy for the iterated prisoner’s dilemma. In: Proceedings of the Fifth International Workshop on the Synthesis and Simulation of Living Systems; 1997. p. 202–209.
  54. 54. Banks JS, Sundaram RK. Repeated games, finite automata, and complexity. Games and Economic Behavior. 1990;2(2):97–117.
  55. 55. Robson AJ. Efficiency in evolutionary games: Darwin, Nash and the secret handshake. Journal of theoretical Biology. 1990;144(3):379–396. pmid:2395377
  56. 56. Unknown.; 2017.
  57. 57. Li J, Kendall G, Member S. The effect of memory size on the evolutionary stability of strategies in iterated prisoner’s dilemma. 2014;X(X):1–8.
  58. 58. Ashlock D, Kim EY. Fingerprinting: Visualization and automatic analysis of prisoner’s dilemma strategies. IEEE Transactions on Evolutionary Computation. 2008;12(5):647–659.
  59. 59. Carvalho AL, Rocha HP, Amaral FT, Guimaraes FG. Iterated Prisoner’s Dilemma-An extended analysis. 2013;.
  60. 60. Adami C, Hintze A. Evolutionary instability of zero-determinant strategies demonstrates that winning is not everything. Nature communications. 2013;4(1):2193. pmid:23903782
  61. 61. Kraines D, Kraines V. Pavlov and the prisoner’s dilemma. Theory and decision. 1989;26(1):47–79.
  62. 62. Kuhn S. Prisoner’s Dilemma. In: Zalta EN, editor. The Stanford Encyclopedia of Philosophy. spring 2017 ed. Metaphysics Research Lab, Stanford University; 2017.