probcomp
Multistate conditional sampling is inaccurate
Conditional draws from each model are given equal weight. This is inaccurate. Models should be drawn from the multinomial given by the total probability of the conditions.
The same should be true for weighting the estimates of logpdf by probability of the conditions under each State.
For reference, gpmcc.Engine implements the weighting as post-processing steps on the entire set of samples and logpdfs returned from each State (which could be improved for simulate in order to remove the dependence between the resampled draws).
https://github.com/probcomp/gpmcc/blob/master/src/engine.py#L243-L256
I'd love to understand this better (problem and solution). May I ask for a fifteen minute tutorial at some point?
On Mon, May 2, 2016 at 12:55 PM, F Saad [email protected] wrote:
The same should be true for weighting the estimates of logpdf by probability of the conditions under each State.
For reference, gpmcc.Engine implements the weighting as post-processing steps on the entire set of samples and logpdfs returned from each State (which could be improved for simulate in order to remove the dependence between the resampled draws).
https://github.com/probcomp/gpmcc/blob/master/src/engine.py#L243-L256
— You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub https://github.com/probcomp/crosscat/issues/97#issuecomment-216291313
Gregory Marton
[email protected] 617-858-0775 http://csail.mit.edu/~gremio/cal
Sure, I'd be happy to. Just let me know when would be good.
On Mon, May 2, 2016 at 11:24 PM, Gregory Marton [email protected] wrote:
I'd love to understand this better (problem and solution). May I ask for a fifteen minute tutorial at some point?
On Mon, May 2, 2016 at 12:55 PM, F Saad [email protected] wrote:
The same should be true for weighting the estimates of logpdf by probability of the conditions under each State.
For reference, gpmcc.Engine implements the weighting as post-processing steps on the entire set of samples and logpdfs returned from each State (which could be improved for simulate in order to remove the dependence between the resampled draws).
https://github.com/probcomp/gpmcc/blob/master/src/engine.py#L243-L256
— You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub https://github.com/probcomp/crosscat/issues/97#issuecomment-216291313
Gregory Marton
[email protected] 617-858-0775 http://csail.mit.edu/~gremio/cal
— You are receiving this because you authored the thread. Reply to this email directly or view it on GitHub https://github.com/probcomp/crosscat/issues/97#issuecomment-216426683