adaptive sampling about dynesty HOT 3 CLOSED

Great idea -- I'll see if I can add this in once I get the other sampling methods put in place.

from dynesty.

Referenced issue incorrectly in (now overwritten) commit.

from dynesty.

After thinking about this a bit more, it's unclear a priori whether one sampling method will be more or less efficient at exploring the prior volume than another; the detailed efficiency properties depend primarily on the dimensionality of the space but also to some extent on the shape of the target distribution. While I could likely assign some rough criteria for switching from one to another (via rough dimensional scalings for 'rwalk', 'slice' and 'rtraj'), I personally think these other sampling methods are more useful when the distribution cannot be effectively bounded. This holds true (1) when the number of live points is small and (2) when the dimensionality of the problem is high (since the volume of bounding surfaces become increasingly concentrated near the boundaries). These seem to match the general framework of the (in progress) DynamicSampler better, so I think this type of adaptive sampling fits much more naturally there.

I guess my thinking is I don't think it's worth adding this feature natively into NestedSampler at the moment. However, given that I'll be putting in the framework to do very similar things with DynamicSampler, I might come back to this if it turns out to be a straightforward addition.

from dynesty.