Hi Christian,
I’ve tried by repeating the recovery for each subject individually, and with centered parametrisation (learning rate ~ Beta(1,1), inverse temperature ~ HalfNormal(sigma=10), decay ~ Beta(1,1)) and things improve a bit.
I also tried by increasing the number of trials from 80 to 200 (maintaining reversals every 20 trials) and the difference is not massive - pearson r for decay goes from 0.52 to 0.62.
I changed the distributions from which I sample the simulation parameters to be more precise (in particular for decay, which I now changed to Beta(0.5,1.5) as most of the decay parameters should be small) and it looks much better (single subject fitting, centered parametrisation, 80 trials).
Based on this, I guess it is an issue with the model not behaving well across the entire parameter space and requiring stronger priors on its parameters.
Could this make sense?
In case, what would be the most sensible way to adapt the Beta(0.5,1.5) to work with a non-centered parametrisation?