search

openforcefield / open-forcefield-tools

Tools for open forcefield development
MIT License
8 stars 6 forks source link

Target uncertainty calculation #12

Closed jchodera closed 8 years ago

jchodera commented 8 years ago

Question for @davidlmobley and @mrshirts:

The current strategy for PropertyEstimator.computeProperties() will be, for each property requested and parameter set:

My question is this: What uncertainty threshold should we use for this, and how much flexibility should we give the user over it?

My original idea was to fix this at, say, 0.1 times the experimental uncertainty (to ensure that our Bayesian likelihood functions are meaningful). Does this sound reasonable, and would you still like control over this (in case you want to make it more or less strict)?

jchodera commented 8 years ago

Also, is this relative uncertainty (based on experimental uncertainty) an OK way to specify this, or do you need more control?

davidlmobley commented 8 years ago

I think that's a great default target, and I think you should allow us to override that. Seems like it would be nice to be able to override by providing either (a) a multiple of the experimental uncertainty to hit, or (b) an absolute target uncertainty - for example, if I was doing density calculations for a big set of compounds and I wanted to be able to report all of them to 1e-5 g/cm^3 for consistency across the set, it'd be annoying to have to specify it for all of them in units of variable experimental uncertainties....

David

On Mon, Jul 4, 2016 at 7:21 PM, John Chodera notifications@github.com wrote:

Also, is this relative uncertainty (based on experimental uncertainty) an OK way to specify this, or do you need more control?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/open-forcefield-group/open-forcefield-tools/issues/12#issuecomment-230374965, or mute the thread https://github.com/notifications/unsubscribe/AGzUYQuz8u0Xh6ewiGnyYAPkglds661Yks5qSb-YgaJpZM4JEuoz .

David Mobley dmobley@gmail.com 949-385-2436

davidlmobley commented 8 years ago

I talked to @mrshirts and he agrees. There is a small concern that some calculations may be sufficiently costly that we may not be able to converge them that precisely in any reasonable amount of time, but we can cross that bridge when we come to it. Marking resolved.