Machine-Learning to Trust

Abstract

Can society sustain long-run mutual trust when agents’ equilibrium beliefs are shaped by machine-learning predictive methods? I study an infinite-horizon game with state-dependent payoffs, in which each player in his turn decides whether to place trust in his immediate successor. Players best-reply to a probabilistic tit-for-tat belief, which is a coarse fit of the true population strategy with respect to a partition of relevant contingencies. In equilibrium, this partition minimises the sum of the mean squared prediction error and a complexity penalty proportional to its size. Relative to symmetric mixed-strategy Nash equilibrium, this solution concept can significantly narrow the scope for mutual trust.