Extended Data Fig. 3: Behavior was better explained by a model that extended S–R learning with state inference than by models that implemented alternative hypotheses.
The winning state-inference model (model S–S–R–8) explained behavior better than models that implemented learning rates that could vary across the task (S–Rα(T) models) or models that implemented stimulus stickiness (S–RStimStick models; see ‘Alternative model families’ in ‘Computational models’ in the Methods and in the Supplementary Methods). a, Model frequencies. b, PEPs. The horizontal red line indicates the threshold for confident selection of a model. State-inference model S–S–R–8 was confidently selected as the best model.
