Async bo rule and notebook #366
Conversation
trieste/acquisition/rule.py
Outdated
| we want to immediately use acquisition function to launch a new observation and avoid wasting computational resources. | ||
|
|
||
| To make the best decision about next point to observe, acquisition function needs to be aware of currently running observations. | ||
| We call such points "pending". AsyncEGO provides a ``AsyncEGO.State`` object that keeps track of pending points |
There was a problem hiding this comment.
I think we should differentiate between "pending" points used by greedy acquisition functions (points that have been selected but haven't been executed yet) and "outstanding" points, which have been executed but haven't returned.
| We call such points "pending". AsyncEGO provides a ``AsyncEGO.State`` object that keeps track of pending points | |
| We call such points "outstanding". AsyncEGO provides a ``AsyncEGO.State`` object that keeps track of outstanding points |
There was a problem hiding this comment.
Can you elaborate a bit what's the difference between pending and outstanding? I don't think I understand.
Once the point was chosen by the acquisition function, there is no difference any more, as I understand it. It could have been chosen 5 calls ago or just now, it is still treated exactly the same - as a point which was selected for observations, but result for which hasn't arrived yet.
There was a problem hiding this comment.
It's like the difference between things in your Amazon basket, and things that you've ordered but haven't arrived yet.
- Greedy acquistiion functions choose points one at a time. The first points are pending until we've selected the entire batch.
- Asyn BO choose some points (greedily) and then ask for them to be evaluated (like normal BO), but then starts another loop before all of them have finished executing. The ones that haven't finished yet are outstanding.
|
Hopefully all the comments are addressed. Noticeable updates:
|
This PR replaces and expands #363.
Here we:
Unit tests, docs and other trimmings to come at a later point, I am mostly looking for feedback at this point. One note about new rule is that, unlike TrustRegion, this one works not with other rules but with function builders directly. Happy to hear suggestions of how to change that.
Update: all the trimmings are now done. The notebook is decorated with many-many comments to make it as accessible as possible, very happy to improve it further if people still find it unclear what's all this multiprocessing stuff is about.