Learning to Cooperate: Learning Networks and the Problem of Altruism

We explore how two populations learn to cooperate with each other in the absence of institutional support. Individuals play iterated prisoner's dilemmas with the other population, but learn about successful strategies from their own population. Our agent-based evolutionary models reconfirm that cooperation can emerge rapidly as long as payoffs provide a selective advantage for nice, retaliatory strategies like tit-for-tat, although attainable levels of cooperation are limited by the persistence of nonretaliatory altruists. Learning processes that adopt the current best response strategy do well only when initial conditions are very favorable to cooperation, while more adaptive learning processes can achieve high levels of cooperation under a wider range of initial conditions. When combined with adaptive learning, populations having larger, better connected learning relationships outperform populations with smaller, less connected ones. Clustered relationships can also enhance cooperation, particularly in these smaller, less connected populations.