Learning Transference Between Dissimilar Symmetric Normal-Form Games

Learning and adaptive models in the economics literature are built on the premise that actions that have yielded higher payoffs in the past are more likely to be selected in the present. Such action-based learning models are not directly applicable to learning that occurs between situations where the set of actions is not directly comparable from one situation to the next. One way to account for the transfer of learning between such situations or games is to relabel actions based on some common properties. In this work, we examine one framework, known as Level-n, for such a purpose, and combine it with two learning dynamics—Experience Weighted Attraction and Rule Learning—to arrive at predictions for a sequence of ten thriceplayed dissimilar games. Using experimental data, we find that when Experience Weighted Attraction is augmented with simple action re-labeling, it performs well in capturing the between-game transference as it affects initial play in each new game. However, only RuleLearning, without the need for re-labeling, captures the ability of players to learn to reason across games.