In 1980 political scientist Robert Axelrod invited game theorists worldwide to submit computer programs to play the Iterated Prisoner's Dilemma (IPD) — a repeated game in which two players simultaneously choose to cooperate (C) or defect (D) each round, with payoffs T > R > P > S (temptation, reward, punishment, sucker). The surprising winner of both tournaments was the simplest entry: Tit-for-Tat, which cooperates on the first move and then copies the opponent's previous action. Axelrod's analysis, published in "The Evolution of Cooperation" (1984), became a landmark in evolutionary biology, economics, and political science, explaining why cooperation can emerge among self-interested agents without central authority.
This simulation runs a full round-robin tournament with 10 strategies, including Tit-for-Tat, Generous TFT, Pavlov (Win-Stay Lose-Shift), Grudger, Random, and Always Defect. You can set the number of rounds per match, customise the TRPS payoff matrix, add noise (ε % chance of accidental defection), and watch the score matrix and cooperation rates evolve live. Noise reveals which strategies are robust to mistakes — plain TFT spirals into mutual punishment, while Generous TFT and Pavlov can escape.
What is the Prisoner's Dilemma?
Two suspects are interrogated separately. Each can betray the other (Defect) or stay silent (Cooperate). If both stay silent, both get a light sentence (reward R). If one betrays and the other stays silent, the betrayer goes free (temptation T) while the other gets the maximum sentence (sucker's payoff S). If both betray, both get a moderate sentence (punishment P). The payoff ordering T > R > P > S makes defection the dominant strategy in a single game — yet mutual cooperation gives a better outcome for both than mutual defection.
Why did Tit-for-Tat win Axelrod's tournaments?
Axelrod identified four properties that made TFT successful: it is "nice" (never defects first), "retaliatory" (immediately punishes defection), "forgiving" (returns to cooperation after one round of retaliation), and "clear" (its strategy is obvious to the opponent, enabling coordination). These properties let TFT do very well against cooperative partners (gaining R per round) while limiting damage against defectors (losing S only once per encounter). No single strategy beats TFT in all environments, but TFT accumulates the most total points across diverse opponents.
What does noise (ε) do to the tournament?
Noise models accidental defection: with probability ε each intended cooperation is flipped to a defection. For plain Tit-for-Tat, a single noise event triggers a retaliatory defection, which triggers another, locking both players into mutual punishment until the match ends — a phenomenon called "echo defection". Generous TFT (which cooperates with probability 1 – ε when retaliating) and Pavlov (Win-Stay Lose-Shift) can escape this trap by occasionally forgiving, making them far more robust in noisy environments.
Pavlov cooperates if both players made the same move last round (both C or both D — a "win") and defects if they made different moves (one C one D — a "loss"). This simple reinforcement rule is self-correcting: after a noise-induced double-defection (punishment P), Pavlov defects while TFT cooperates — switching back to mutual cooperation after one round. Pavlov can also exploit Always-Cooperate by switching to permanent defection, giving it an advantage TFT lacks. In noisy IPD tournaments Pavlov often outperforms TFT.
Generous TFT (GTFT) is like Tit-for-Tat but instead of always retaliating after an opponent's defection, it cooperates with a small probability p (typically p = 1/3 or matched to the noise rate). This forgiveness breaks echo-defection cycles: after a chain of mutual punishment, GTFT occasionally cooperates, giving the opponent a chance to return to mutual cooperation. Nowak and May (1992) showed that GTFT is an evolutionarily stable strategy in spatially structured populations, displacing both TFT and Always-Defect.
Grudger cooperates on every move until the opponent defects even once, then defects forever after — holding a permanent grudge. Against cooperative opponents, Grudger accumulates the maximum reward. Against any defector, after the initial sucker loss it switches to mutual punishment, avoiding further exploitation. While Grudger is not "forgiving" and performs poorly against noisy or testing opponents, it is an equilibrium concept: once both players grudge-trigger, neither can benefit by unilaterally switching back to cooperation.
Always Defect (AllD) earns the temptation T when facing cooperators, which seems profitable. But in a round-robin tournament it earns only the punishment P (the worst mutual outcome) against every other defecting strategy — and against any version of TFT or Grudger it gains T only on the first move, then collects P for the remaining N – 1 rounds. Against many cooperative opponents, AllD's aggregate score is crushed because R >> (T + (N-1)P)/N for large N. AllD only "wins" environments composed mainly of naive cooperators.
The payoff matrix must satisfy T > R > P > S and 2R > T + S for the iterated dilemma to be meaningful. Increasing T (temptation) raises the reward for defecting against cooperators, making the environment harsher for nice strategies. Decreasing the gap between R and P narrows the incentive to cooperate. When T – R becomes very large, even TFT begins to defect opportunistically in some variants. Adjusting the sliders lets you explore how the strength of the dilemma affects which strategies dominate.
Robert Axelrod and William Hamilton's 1981 paper showed that cooperation can invade a population of defectors if the discount factor w (probability of another round) satisfies w ≥ (T – R)/(T – P). Above this threshold the future shadow of repeated interaction makes defection unprofitable. Kin selection (Hamilton's rule: rb > c) and group selection provide additional evolutionary pathways. In practice, spatial structure — where cooperators can form clusters and avoid exploiters — is often the key mechanism, as shown by Nowak and May's spatial IPD lattice models.
In a round-robin tournament every participant plays every other participant (and, in Axelrod's setup, also themselves) for N rounds. The total score is the sum of payoffs accumulated across all matches. Scores are therefore influenced not just by how well a strategy does against defectors but also by how well it does against the majority of cooperative strategies in the pool. This is why the composition of the field matters: AllD performs better in a defection-heavy field, while TFT thrives in a cooperation-heavy one.
The IPD models many social and biological phenomena: arms races between nations (cooperate = limit weapons, defect = escalate); corporate price competition (cooperate = hold price, defect = undercut); immune evasion by pathogens (cooperate = suppress toxin, defect = burst); cleaner fish and their hosts; even synaptic plasticity rules in neural networks. Axelrod's insights have been used to design international trade agreements, suggest mechanisms for climate cooperation, and explain the evolution of altruism in social insects.