Run longer to get better answers?

14 views
Skip to first unread message

Tom Lieber

unread,
Jan 14, 2024, 9:52:50 AM1/14/24
to webppl-dev
I'm not familiar enough with the inference algorithms to choose good parameters yet, but I do know how long I'm willing to wait for an answer. Is there a way to run for a particular amount of time instead of for a particular number of steps?

Related to that, is there a way to run inference such that it picks up where it left off? For example, run MCMC for 100 steps, print the result of a query, run for 100 steps more, print improved results, etc.

null-a

unread,
Jan 17, 2024, 10:19:46 AM1/17/24
to webppl-dev
Hi. It's been a while since I worked with WebPPL, but fwiw, as best I remember...

> Is there a way to run for a particular amount of time instead of for a particular number of steps?

In general, I'm afraid not. (Though I think inference by enumeration might have this as an undocumented feature.)

> Related to that, is there a way to run inference such that it picks up where it left off?

Again, I don't think so.

Best,

Paul

Hakan Kjellerstrand

unread,
Jan 17, 2024, 11:12:33 AM1/17/24
to null-a, webppl-dev
On Wed, Jan 17, 2024 at 4:19 PM null-a <hors...@gmail.com> wrote:
Hi. It's been a while since I worked with WebPPL, but fwiw, as best I remember...

> Is there a way to run for a particular amount of time instead of for a particular number of steps?

In general, I'm afraid not. (Though I think inference by enumeration might have this as an undocumented feature.)

The feature is maxRuntimeInMS (where MS=milliseconds), and is only supported for method="enumerate".
Example:
    var d = Infer({method:"enumerate",maxRuntimeInMS:6000000},model)

Best,

Hakan

On Sunday, January 14, 2024 at 2:52:50 PM UTC all...@gmail.com wrote:
I'm not familiar enough with the inference algorithms to choose good parameters yet, but I do know how long I'm willing to wait for an answer. Is there a way to run for a particular amount of time instead of for a particular number of steps?

Related to that, is there a way to run inference such that it picks up where it left off? For example, run MCMC for 100 steps, print the result of a query, run for 100 steps more, print improved results, etc.

--
You received this message because you are subscribed to the Google Groups "webppl-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to webppl-dev+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/webppl-dev/4c43ee27-030b-45c9-805f-ae486f1acbf6n%40googlegroups.com.


--

Tom Lieber

unread,
Jan 30, 2024, 1:20:23 PM1/30/24
to webppl-dev
Thanks, all.
Reply all
Reply to author
Forward
0 new messages