Fig. 6: Experiment 1 model recoverability.
From: Reinforcement learning increasingly relates to memory specificity from childhood to adulthood
For each model within each stage of model comparison, 100 simulated āexperimentsā were conducted in which choice data were simulated from 151 agents, with parameters sampled from uniform distributions with ranges determined by the empirical fits. Data from each simulated experiment were then fit with each model within the comparison set. The top panels show confusion matrices, where the values within each tile represent the proportion of experiments for which each fitted model had the highest exceedance probability (top panels). The bottom panels show inversion matrices, where the values within each tile represent the proportion of experiments for which the fitted model had the highest exceedance probability that were generated by each of the models. Black lines outline the model that best fits the empirical data within each comparison stage. A Models with different numbers of choice weights were highly distinguishable from one another. B Models in which exemplar and category values were initialized with either one or two free parameters were distinguishable from a model in which exemplar and category values were both initialized at 0. However, models with one or two initial values were not distinguishable from one another. C The winning model, in which participants learned equivalently from experienced and counterfactual outcomes, was highly distinguishable from a model in which participants learned with separate learning rates for experienced and counterfactual outcomes, as well as a model in which participants did not learn from counterfactual outcomes.
