For analyses of data simulated from different RL agent types (Figure 2), we first fitted each agent to our baseline behavioral dataset using the hierarchical framework outlined above. The agents used were a model-free agent with eligibility traces and value forgetting (Figures 2D–2F), and a model-based agent with value and transition probability forgetting (Figures 2G–2I) and the best fitting RL model detailed in Figure S3 (Figures 2J–2L). We then simulated data (4000 sessions each of 500 trials) from each agent, drawing parameters for each session from the fitted population level distributions for that agent. We performed the logistic regression on the simulated data, using the same hierarchical framework as for the experimental data.
