Adaptive Directional Learning in Games

Abstract

In this paper, we develop a novel learning model that combines directional learning and automata reinforcement. We show that directional learning can be defined on a cyclic order in addition to its original domain (interval) and that pattern learning by reinforcing parts of automata [Chernov, 2020] can be used with directions instead of actions as a basis of automata. We test our model on two laboratory experimental datasets (Oligopoly [Duersch et al., 2010] and Rock-Paper-Scissors [Chernov, 2020]). Fitting participants’ trajectories and aggregating areas where players’ average payoffs converge, we establish that our model outperforms classical models (Fictitious Play and Reinforcement Learning) and produces a behavior aligned with human decisions while making fewer mistakes.

Keywords: Directional Learning, Cross-Evaluation, Repeated Games