Action-Reinforcement Learning Versus Rule Learning

How well do various learning models predict the dynamics of the population distribution of play in a variety of games? We measure and compare the in-sample and out-of-sample prediction performance of seven action-reinforcement learning models as well as Rule Learning for symmetric normal-form games. The criteria include log-likelihood values and root-mean-squared error. We conclude that a logit best-reply dynamic with inertia and adaptive expectations (LBRIAE) is the best among the action-reinforcement models. We formulate a population Rule Learning model that nests the LBRIAE model and find that Rule Learning is statistically superior. Sophistication is learned in the sense that random behavior and herd behavior decline substantially, and level-1 and level-2 behaviors increase over time