Evaluating generalizability and parameter consistency in learning models

A new evaluation method is proposed for comparing learning models used for predicting decisions based on experience. The method is based on the generalization of models' predictions at the individual level. First, it evaluates the ability to make a priori predictions for decisions in new tasks using parameters from different tasks performed by an individual decision-maker. Second, it evaluates the consistency of parameters estimated in different tasks performed by the same person. We use this method to examine two rules for updating past experience with payoff feedback: The Delta rule, where only the chosen option is updated; and a Decay-Reinforcement rule, where additionally, non-chosen options are discounted. The results reveal that although the Decay-Reinforcement rule fits the data better, it has poor generality and parameter consistency at the individual level. The current method thus improves the ability to select models based on their correspondence to consistent characteristics within individual decision-makers.