MCSim simulation sofware - Tasks: task #15666, improve tempered MCMC
You are not allowed to post comments on this tracker with your current authentication level.
task #15666: improve tempered MCMC
Submitter: | Frederic Yves Bois <fredomatic> | ||
Submitted: | Thu 28 May 2020 01:31:10 PM UTC | ||
Should Start On: | Wed 27 May 2020 10:00:00 PM UTC | Should be Finished on: | Thu 27 May 2021 10:00:00 PM UTC |
Category: | Sim | Priority: | 5 - Normal |
Status: | None | Assigned to: | None |
Percent Complete: | 0% | Open/Closed: | Open |
Attached Files
Depends on the following items: None found
Items that depend on this one: None found
Carbon-Copy List
Follow 2 latest changes.
Date | Changed by | Updated Field | Previous Value | => | Replaced by |
---|---|---|---|---|---|
2020-05-28 | fredomatic | Attached File | - | Added linear.model, #49184 | |
Attached File | - | Added linear.LTMCMC.in, #49185 |
Powered by Savane 3.13-f8d8.
Corresponding source code
Jumping between perks can deteriorate when adding a new perk because if the new perk attempted pseudo-prior is wrong, Robbins-Munro updating can disrupt the pseudo-prior of the perk just above the new one. The bottom perks get cut off from the higher perks. The damage is almost irreversible, because slopes of the pseudo-prior vs perk curve go bad and are used to set the pseudo-prior of newly attempted points.
This can be seen with the hierarchical linear model (attached).
Maybe the "old good" set of pseudo-priors should be saved and re-instated when trying a new perks when the previous attempt has failed.
Another option is to go from 1 down to 0.5 (or 0.1) and also from 0 up to 0.5 (or 0.1) and then bridge the two.
Maybe there should be no more that 0.2 or 0.3 between two perks. That does not fix the above problem anyway.
Decent tests should be set up to check that fixing that problem does not botch what works already.