A Look at Learning Models in Three-by-Three Bimatrix Games

Experimental data is used to test a variety of learning models using a model that extends several of the existing learning models. Generally, the parameter estimate are in the expected ranges. Individual agent parameter estimates indicate that there is considerable individual heterogeneity. Representative agent parameter estimates adequately predict the mode of the individual parameter estimates when the data is pooled across matrices. They are not very effective at predicting the mode of the disaggregated data. There is some evidence in favour of the restriction that the two discounts are equal. The restrictions that the agents equally weigh actions experienced and actions not experienced is rejected using both representative agent and individual agent parameter estimates) thought there is evidence that subjects put more weight on actions experienced over those not experienced). We reject the rote learning version of the model in favour of a (weak) belief learning parameterizations of the model.