Selection of learning rules: theory and experimental evidence

David J. Cooper and Nick Feltovich

The goal of this paper is to formulate and apply a rule for mapping between experimental designs and learning rules. We propose the use of a hierarchy of learning rules, with rules higher on the hierarchy possessing a greater degree of cognitive sophistication, and suggest that the learning model with the least cognitive sophistication should be used unless certain well-specified criteria are met for moving to a more sophisticated model. These ideas are applied to data from limit pricing experiments. We compare the abilities to characterize the data of a reinforcement-based learning model and a belief-based learning model, and find that the belief-based model outperforms the reinforcement-based model. This is not due to some universal superiority of belief-based learning models. Rather, the belief-based model's greater cognitive sophistication makes it the appropriate model for the limit pricing data.

Paper (PDF)