worry about likelihoods underflowing in SMCTC about sts

[koadman]

well spotted. i think we care only about creating a sampling distribution from the weights and not about the actual likelihoods.
if that's true it seems like we should be able to rescale and still get the desired result. pseudocode follows:

m = max(log_weights)
log_weights += m

some very low likelihoods may still underflow with the exp() but those losers weren't going to make it to the next generation anyway. hehehe.

[matsen]

Hrm, that ain't going to work for stratified resampling, if you see #27. We might be able to make it work in log-space but it will require some thought. Even multinomial calls

double GetWeight(void) const {return exp(logweight);}

... but there it seems straightforward to first subtract off the smallest log like then exponentiate, then take the multinomial. Seem good????

[koadman]

ja, i looked over that smctc patch, seems ok. looks like you were able to get rid of a few extra calls to exp() as well. nice!

[matsen]

So you think that using the first likelihood as a normalizing factor is an
OK choice? Seemed better to me than the most likely, and faster.

On Sun, Oct 21, 2012 at 9:07 PM, Aaron Darling [email protected]:

ja, i looked over that smctc patch, seems ok. looks like you were able to
get rid of a few extra calls to exp() as well. nice!

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/30#issuecomment-9652128.

Frederick "Erick" Matsen, Assistant Member
Fred Hutchinson Cancer Research Center
http://matsen.fhcrc.org/

[koadman]

well, i guess it depends how much range of variation there is in likelihoods. if they vary over more than e.g. 300 log units then the high likelihoods would overflow. Seems like it's much more important to keep the high likelihoods intact than the little'uns. Why did you not want to normalize with the most likely?

[matsen]

Good point.

I guess I was thinking that on average we would have twice as much range if
we took something that was in the middle rather than something at the top.

I just had a little play around and noted that exp(-big) = 0 and exp(+big)
= inf.

So we would certainly know either way.

However, I agree and will fix and then merge.

On Mon, Oct 22, 2012 at 7:30 AM, Aaron Darling [email protected]:

well, i guess it depends how much range of variation there is in
likelihoods. if they vary over more than e.g. 300 log units then the high
likelihoods would overflow. Seems like it's much more important to keep the
high likelihoods intact than the little'uns. Why did you not want to
normalize with the most likely?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/30#issuecomment-9665733.

Frederick "Erick" Matsen, Assistant Member
Fred Hutchinson Cancer Research Center
http://matsen.fhcrc.org/

[koadman]

if we want the extra range maybe we could find the max particle weight
and subtract off 300 or so log units from it (assuming double precision
floats) to get the normalization weight.

On Mon, 2012-10-22 at 08:31 -0700, Erick Matsen wrote:

Good point.

I guess I was thinking that on average we would have twice as much
range if
we took something that was in the middle rather than something at the
top.

I just had a little play around and noted that exp(-big) = 0 and
exp(+big)
= inf.

So we would certainly know either way.

However, I agree and will fix and then merge.

On Mon, Oct 22, 2012 at 7:30 AM, Aaron Darling
[email protected]:

well, i guess it depends how much range of variation there is in
likelihoods. if they vary over more than e.g. 300 log units then the
high
likelihoods would overflow. Seems like it's much more important to
keep the
high likelihoods intact than the little'uns. Why did you not want
to
normalize with the most likely?

—
Reply to this email directly or view it on
GitHubhttps://github.com//issues/30#issuecomment-9665733.

[cmccoy]

Closed - just dividing by the largest log-like for now. Anyone feel strongly that we need more range than that gives?