We report the results of a coordination game experiment. The experiment carefully distinguishes between conventions based on labels and conventions based on populations. Our labels treatments investigate the abstraction assumptions that underlie the concept of a strategy, while our population treatments investigate the attraction of alternative mutually consistent ways to play under adaptive behaviour. We observe conventions emerging in communities with one population and labels and with two populations and no labels, but the most effective treatment is two labelled populations. We estimate logistic response learning models for individual subject behaviour. Of the models considered, a version of exponential fictitious play fits our data best.