The evolution of inefficiency in a simulated stag hunt

We used genetic algorithms to evolve populations of reinforcement learning (Q-learning) agents to play a repeated two-player symmetric coordination game under different risk conditions and found that evolution steered our simulated populations to the Pareto inefficient equilibrium under high-risk conditions and to the Pareto efficient equilibrium under low-risk conditions. Greater degrees of forgiveness and temporal discounting of future returns emerged in populations playing the low-risk game. Results demonstrate the utility of simulation to evolutionary psychology.