Axelrod's Prisoner's Dilemma tournaments

Axelrod’s Prisoner’s Dilemma tournaments tested how cooperation fares against defection when strategies meet repeatedly. Their central result was not that kindness always wins, but that a simple, initially cooperative strategy that responds to an opponent’s behavior can prosper in the right strategic environment; an evolutionary follow-up further showed how the composition of that environment changes over time.

The game and the tournament design

In the human version of the Prisoner’s Dilemma, two players each choose, without seeing the other’s choice, whether to cooperate or defect. A banker reveals both cards and pays out according to the resulting combination. Because each player’s reward depends on both choices, the game makes individual incentives and mutual outcomes visibly distinct. The tournaments translated this structure into repeated play, where a strategy was a preprogrammed rule for choosing actions across successive rounds.

For the first competition, Axelrod invited experts in game theory to submit strategies in computer language. Fourteen were entered, and Axelrod added a fifteenth, Random, which selected cooperation or defection at random as a baseline. He translated all 15 into a common programming language and arranged every strategy to play each other strategy, including a copy of itself. This produced 225 separate games. Each pairing played 200 moves; the resulting winnings were added across a strategy’s 15 pairings to determine its overall score. The scoring system assigned three points to mutual cooperation, five to the temptation to defect, one to mutual defection, and zero to the sucker’s payoff.

Tit for Tat and the second competition

The first tournament’s winner was Tit for Tat, submitted by Anatol Rapoport. Its rule was spare: cooperate on the first move, then copy the other player’s previous move. Against another copy of itself, the strategy sustained cooperation throughout the game, allowing both players to receive the full benchmark score of 600 points. The result made a simple rule look unexpectedly powerful beside more elaborate submissions.

Axelrod’s second tournament drew 62 entries, with Random again added, for 63 strategies altogether. Unlike the first competition, the number of moves in each game was left open rather than fixed at 200. Entrants had the first tournament’s results and Axelrod’s explanation of why nice and forgiving strategies had done well. Some responded by submitting strategies of that kind; others expected their competitors to do the same and entered strategies intended to exploit them. John Maynard Smith submitted the forgiving Tit for Two Tats.

Tit for Tat won again, scoring 96 percent of the benchmark. Nice strategies generally ranked well: all but one of the top 15 were nice, while all but one of the bottom 15 were nasty. But Tit for Two Tats, which would have won the first tournament had it been entered, did not win the second. Its unusually forgiving rule could be exploited by more subtle nasty strategies. The contrast showed that a strategy’s tournament success depended partly on which rivals happened to be present, not on its rule in isolation.

From competition to evolutionary change

To move beyond the arbitrary composition of a submitted field, Axelrod used the same 63 strategies in a computer simulation treated as an evolutionary succession. Generation 1 began with equal representation of every strategy. At the end of each generation, winnings became offspring identical to their asexual parents; as a result, strategies that did well grew more numerous, while others became scarce or disappeared. Their changing proportions altered the strategic climate in which later generations played.

After about 1,000 generations, the proportions stopped changing. Most strategies had already gone extinct by generation 200, though some nasty strategies initially increased. Harrington was the only nasty strategy to survive beyond that point: its frequency rose steeply for about 150 generations, then gradually declined toward extinction around generation 1,000. Its temporary success came from exploiting overly forgiving strategies such as Tit for Two Tats. Once those easy targets had been driven out, Harrington lost its advantage, leaving room for nice but provocable strategies.

Tit for Tat finished first in five of the six runs, and five other nice but provocable strategies ended nearly as successful; one of them won the remaining run. When nasty strategies disappeared, the surviving nice strategies cooperated with one another and became indistinguishable in practice. This result qualified the apparent triumph of Tit for Tat: it could resist invasion by nasty strategies, but it was not strictly an evolutionarily stable strategy, because a different nice strategy could enter a population of nice cooperators without being detected or gaining an immediate disadvantage. That distinction connected the tournament’s findings to the broader idea that a strategy’s success must be judged in a population where it is already common, not only against a hand-picked field of competitors.

Cooperation beyond the computer

The tournaments illustrate how repeated interactions can support cooperation without requiring conscious foresight in every participant. The book extends this analytical framework from computer strategies to life, describing organisms as following preprogrammed rules in the evolutionary sense. In that account, cooperative conduct can serve the longer-term success of genes even though the underlying evolutionary process is not purposeful. The explanation therefore treats apparent strategic intent as a useful metaphor, not evidence that genes or programmed strategies consciously plan.

The same distinction between rules and incentives informs the book’s discussion of human arrangements. Divorce might allow a separating couple to cooperate for mutual benefit, including by reducing legal expense and preserving resources for children. But in England and, until recently, in all fifty states of the United States, professional rules required a lawyer to accept only one member of a couple as a client. That structure could cast the dispute in adversarial terms even if the couple preferred an amicable settlement, with the possibility that the lawyers benefited from the resulting contest. Rules do not merely administer a conflict; they can help determine whether participants encounter it as a zero-sum struggle or a situation with scope for mutual gain.

A contrasting example came from English league football. On May 18, 1977, Bristol and Coventry were playing the match that would decide whether either faced relegation. Sunderland’s result mattered to both: once news arrived that Sunderland had lost, a draw was enough for Bristol and Coventry to remain in the First Division. The Bristol-Coventry match was five minutes behind schedule, so the result became known before it ended. After the news was displayed at the ground, both teams had reason to preserve the draw rather than risk a goal that could leave one of them relegated. What had been a contest between rivals thus became a situation in which both could benefit from the same outcome.

The book uses such examples to emphasize the importance of repeated play and uncertain endings. If players know exactly when their interaction will stop, the final round can invite defection; reasoning backward can unravel cooperation in earlier rounds as well. In an indefinitely or unpredictably repeated game, by contrast, the prospect of future encounters can sustain reciprocity. The tournaments’ durable lesson is therefore conditional: cooperation can flourish within a basically competitive world when strategies, repeated encounters, and surrounding rules make it viable, but it is not guaranteed by goodwill alone.