Throwback

Open michito744 opened this issue 1 year ago • 3 comments

There is a example of cases where a wrong conclusion is reached by choosing a clearly ludicrous progression in 28b, when it could resolve in 18b.

#796

2024-06-29 2024-06-29 (2)

This is the effect of 28b's clearly incorrect conclusion reflected in the policy, since KataGo itself discovers the correct conclusion when it is allowed to search deeper in a slightly further phase.

Jun 29 '24 04:06 michito744

Thanks! As with the previous cases, can you attach an SGF?

Jun 29 '24 13:06 lightvector

@lightvector

I placed the SGF in https://github.com/lightvector/KataGo/issues/796 .

Jun 30 '24 03:06 michito744

Ah, thanks! I didn't realize this was the same example as another from before.

Jul 01 '24 12:07 lightvector

@lightvector

The “slowness” inherent to 28b is still a problem, but there has been a clear improvement in this issue. I'll close this page for now. 2024-08-27 2024-08-27 (1)

Aug 27 '24 14:08 michito744