Adaptive learning models that have been tested against experimental data typically share two features: (i) initial attractions (or beliefs) are given exogenously, and (ii) learning is based on the performance of stage-game actions rather than repeated game strategies. We develop a model of learning which endogenizes initial attractions and allows for the learning of repeated game strategies. Learning occurs in two phases. In an initial long-run `pre-experimental' phase, we allow players to explore a complete set of repeated game strategies that satisfy a complexity constraint. The limiting attractions from the first phase are then used as initial attractions in the second, short-run phase, which can be tested against experimental data. We find that, relative to existing adaptive models, we are better able to account for the behavior of subjects in environments where fairness and reciprocity appear to play a significant role.