Implement parameter recovery simulations

[drbenvincent]

In order to have faith that everything is working, we need to be able to perform some parameter recovery simulations.

The goal

Be able to reproduce figures like this

Basic functionality

• [x] implement get_simulated_response() properly
• [x] add vertical line of true parameters onto histogram plots

Technical debt to clear

• [x] Currently we have darc-specific information on how to convert between the Design named tuple and the Pandas DataFrame. This is get_simulated_response(). We need to resolve this so that everything in bad is entirely clueless about anything darc related.

Parameter recovery

• [x] add a parameter recovery function
• [x] make it work
• [x] return summary info from estimated parameters that will be useful to plot
• [x] add parameter sweep
• [x] add ability to plot
• [x] move this into /dev folder as this is not really required or people, it's more for me to check stuff
• [x] add plotting of CI's to the plots

[drbenvincent]

current results: We are getting some discrepant, and unexpected behaviours that deviate from the Matlab results. Todo:

• [ ] check robustness of Kirby
• [ ] why is Kofarnus & Bickel not working as well as it should?
• [ ] Griskevicius is new, but is it really this bad?
• [ ] Frye is acting as expected
• [ ] Our BAD approach: 95% CI's are broader than expected. This is probably happening because of the variation in how we implement the heuristic procedures as compared to the Matlab implementation. This is going to undergo change soon anyway, so we can expect this to improve.

Check that the Python model code is actually using the fixed value of alpha=2 and not in fact inferring it's values.

[drbenvincent]

Actually, much of the discrepancy was because our prior on logk was waaaaay more broad than it should have been.

[drbenvincent]

This issue is now resolved. We can do parameter recovery and parameter recovery sweeps and get some good outputs. Further improvements to our BAD approach will be dealt with elsewhere :)

[drbenvincent]

For the record, the relatively poor performance of our approach was being caused by the heuristic (random_choice_dimension='DB'). Once removed, the estimates for our approach are back to being very good.