search

university-of-newcastle-research / pmwg

Repository for code for the Samplers Team at UoN
3 stars 3 forks source link

Adaptive epsilon value to achieve acceptance rate goal #39

Closed gjcooper closed 1 year ago

gjcooper commented 3 years ago

Is your feature request related to a problem? Please describe. For certain models/datasets getting a good acceptance rate (>30%) is quite difficult, a nice way to deal with this situation would be to, on the fly, change epsilon to achieve that target rate.

Describe the solution you'd like When a special value for epsilon is set to be something like "adaptive" then the epsilon should change based on the recent acceptance rate. Need to come up with a solution for the window over which acceptance rates are calculated, then within certain thresholds, decrease epsilon if acceptance is too low, and increase epsilon if acceptance is too high.

Additional context Data/model from Aline Bompass and Scott Brown would be a good test case due to the low number of participants (3) and low within-subject variability of "response times".

gjcooper commented 1 year ago

This hopefully has been "fixed" with the pull request from Niek Stevenson #64 - an adaptive epsilon (scaling parameter) that is set for each individual.