Miller and Shettleworth (2007) used an associative model of instrumental choice to explain a confusing pattern of results in the geometry learning literature. Dupuis and Dawson (in press) identified a structural flaw in the Miller-Shettleworth (MS) model and suggested replacing it with an operant perceptron model which can correctly reproduce some experimental results that the MS model does not. Here we demonstrate that the error in the MS model can be easily corrected without altering any of the model's predictions by making it stochastic rather than deterministic. In addition, we show that the raw outputs of the perceptron model cannot be interpreted as discriminative choices in an instrumental task without first being normalized. We show that this additional step renders the results of the perceptron model identical to those of the MS model in exactly those cases in which it has been claimed to correctly predict results that the latter cannot.