Home

The Evolution of Strategies in the Iterated Prisoner's Dilemma

Summary of: The Evolution of Strategies in the Iterated Prisoner's Dilemma

Author(s) / Editor(s)

Axelrod, Robert
The genetic algorithm uses computer simulations to evolve different strategies for playing Prisoner's Dilemma games, and by observing the interactions of populations of agents over many runs, it is possible to make useful observations that could generalize to human behavior – such as the tendency of reciprocation to establish itself and spread if cooperating agents are able to encounter one another.

Disciplines

Biology, Computer Science, Economics, Political Science, Information

Keywords

agent-based model, complexity, evolution, game theory, prisoners dilemma, reciprocity, tit-for-tat

Publication Reference

Published in/by
-
Date
1987

Findings

  • Genetic algorithms, developed for complexity and artificial life research, can be used to evolve strategies for playing Prisoner's Dilemma games that are well-adapted to different environments, and thus can be a probe of possible dynamics of human cooperation.
  • From a random start, populations of Prisoner's Dilemma strategies evolve away from cooperation to less cooperative rules, but after a number of runs, those players that reciprocate when encountering cooperation lock into mutually beneficial reciprocal cooperation: reciprocity, once established, can spread through a population that is originally dominated by non-cooperative strategies.
  • Genetic algorithms are highly effective method of searching for successful strategies in very large possibility spaces.

John Holland at University of Michigan developed a means of testing computer problem-solving methods by applying a method based on Darwinian evolution: agents (program) have a phenotype (the strategy the program uses for problem solving) and a genotype (the way strategies are represented in their programming code). Means of reproduction and mutation are specified. Agents interact with each other in a rigorously specified simulation, and the effectiveness of each agent is evaluated in a particular environment in relation to its interactions with other agents; successful strategies are reproduced at a higher rate than less successful strategies; pairs of successful offspring strategies are mated by combining genetic material; mutation is introduced. Simulations can be halted after specified numbers of runs and analyzed, then restarted. In about a quarter of simulation runs with sexual reproduction, better strategies than Tit-for-Tat evolved, and after a random start, populations tend to first evolve away from cooperation as less cooperative rules succeed more often, but can evolve back toward stable cooperation states if cooperative strategies encounter one another and reciprocate.