Fig. 4.

(A) Outcome of two RL agents in the steal/punish game. As the cost of punishing increases, victims become less likely to learn to punish (i.e., become relatively more vulnerable). (B) We embedded the RL agents in an evolutionary simulation, allowing selection for hedonic biases for or against stealing/punishing. As the cost of punishing increases, selection increasingly favors an intrinsic hedonic bias for punishing (rather than stealing) (see Materials and Methods). Prob., probability.