Run multiple dispersed chains by default

[fonnesbeck]

So that convergence can adequately be evaluated, as well as to increase sampling efficiency, the default behavior of sample should be to use at least 2 chains. The starting points for these chains should be overdispersed. One approach could be to run ADVI, and then use the resulting SD to randomly disperse inits from the mean.

A nice extension of this would be to monitor the Gelman-Rubin diagnostic (i.e. Rhat) on the fly, and only begin recording values to the trace once convergence is achieved. This would eliminate the need for manual burn-in.

[aloctavodia]

#1655 is an attempt to provide overdispersed starting points. El ene 11, 2017 5:17 PM, "Chris Fonnesbeck" <notifications@github.com> escribió:

…

So that convergence can adequately be evaluated, as well as to increase sampling efficiency, the default behavior of sample should be to use at least 2 chains. The starting points for these chains should be overdispersed. One approach could be to run ADVI, and then use the resulting SD to randomly disperse inits from the mean. A nice extension of this would be to monitor the Gelman-Rubin diagnostic (i.e. Rhat) on the fly, and only begin recording values to the trace once convergence is achieved. This would eliminate the need for manual burn-in. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1662>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABRuTnEckb_DgIV1S3qAUUWRIL053C3Dks5rRTjrgaJpZM4LhBVe> .

[fonnesbeck]

In #1655, you select random points, correct? To facilitate convergence diagnostics, it may be preferable to have overdispersed inits. For example, a simple approach might be to take, say, points evenly spaced along the circle (or hypersphere, more generally) of 2sd from the optimal point from ADVI/MAP or wherever.

[aloctavodia]

Intuitively I think random initialization is a good idea, but I see your point. What was the default behaviour of PyMC2? what about Stan? El ene 11, 2017 5:45 PM, "Chris Fonnesbeck" <notifications@github.com> escribió:

…

In #1655 <#1655>, you select *random* points, correct? To facilitate convergence diagnostics, it may be preferable to have overdispersed inits. For example, a simple approach might be to take, say, points evenly spaced along the circle (or hypersphere, more generally) of 2sd from the optimal point from ADVI/MAP or wherever. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1662 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABRuTi3f-pGoh-JezVF80gmjgYGDaZOAks5rRT9ngaJpZM4LhBVe> .

[fonnesbeck]

Random is fine for a single chain, but if you are wanting to diagnose convergence, dispersed is more reliable.

PyMC2 uses a random sample (but we were not thinking about multiple chains or parallel sampling back then); Stan uses dispersed, AFAIK.

[fonnesbeck]

I'm going to take a stab at a PR for this. There are no good default starting values for multiple chains. I will use @aloctavodia 's PR as a starting point.

[fonnesbeck]

Just noticed #1655 got merged. Good stuff.