Skip to content
This repository has been archived by the owner on Mar 11, 2021. It is now read-only.

How strong is the model in kyu/dan? #980

Closed
smolendawid opened this issue Mar 18, 2020 · 7 comments
Closed

How strong is the model in kyu/dan? #980

smolendawid opened this issue Mar 18, 2020 · 7 comments

Comments

@smolendawid
Copy link

Excuse me if I can find this info somewhere but I wasn't able to do it. What level for the current model is estimated?

@sethtroisi
Copy link
Contributor

A brief answer is

A lot of strength comes down to number of playouts (which depends on time / hardware).

With even modest hardware I suspect we are 3p+ and on good hardware all the strong engines seem to be able to win regularly against 7p+

@amj
Copy link
Contributor

amj commented Mar 18, 2020

@sethtroisi 3p/7p are no longer indications of strength; they are measures of achievement.

@smolendawid We're definitely past pro level in match settings. A github repo at one point tracked AI strength: https://github.com/breakwa11/GoAIRatings but it doesn't appear to have been updated lately.

@amj amj closed this as completed Mar 18, 2020
@smolendawid
Copy link
Author

smolendawid commented Mar 19, 2020

Great to know, congratulations to all the team. Thanks for the answer

@smolendawid
Copy link
Author

@sethtroisi by a number of playouts you mean --num_readouts? I'd like to play against a strong model but 000737-fury with 10 readouts doesn't seem too strong.

@tommadams
Copy link
Contributor

10 readouts is far too low. You should try at least 800 or 1600

@smolendawid
Copy link
Author

And do you know how many readouts is possible on Raspberry PI in some reasonable time, let's say 5-10 sec?

@tommadams
Copy link
Contributor

I don't know about the full model. But we trained a quantized model that can be run on an Edge TPU, which is very fast: https://coral.ai/projects/minigo/

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

4 participants