.MAP().val

27 views

Skip to first unread message

Claus Möbus

unread,

Jan 16, 2021, 12:25:43 PM1/16/21

to webppl-dev

Dear all,

I found in the BDA-chapter of the eBook an undocumented MaximumumAP-function. I plugged it into the 'observer'model (same chapter). But results seem to be not correct as you can see in this screen-shot:

null-a

unread,

Jan 16, 2021, 1:42:39 PM1/16/21

to webppl-dev

Hi.

It looks like MAP was undocumented because it's not clear what its semantics are, and it might even behave differently depending on which inference algorithm is used to generate the posterior. This old issue is reporting something very similar to what you have encountered:

https://github.com/probmods/webppl/issues/897

Also see:

https://github.com/probmods/webppl/issues/140

In your case, it might be that `posteriorBeliefs` is a bag of samples, each of which has a log prob of 0, and `MAP` is doing argmax over that. If that's not it, I'd guess it's something similar.