lightvector
Best network
As indicated by the ratings, kata1-b60c320-s6321537280-d2951683615 is a very well balanced and powerful network. It withstands very deep searches better than any network in the past, with more accurate predictions in PvP. It would be best to do larger searches on this network and feed back towards upstream.
*Even KataGo's best networks are clearly inferior to FineArt. (Hundreds of thousands more searches are needed to get the same results with KataGo as with FineArt for evaluation).
Thanks. Is this in response to the recent ratings of the different networks, in particular the network after it seeming to be much worse?
I didn't realize at first from your message why you were singling out this particular network, since it looks very similar to many of the networks before it, but it does seem the latest one might be worse, so sure, I went ahead and manually disabled the newest one from training data so it's back to using b60c320-s6.32G.
@lightvector
In various complex phases, it is quite a bit of luck whether the search can be developed in a balanced manner.
Up to s6.32, that accuracy tended to increase, but with s6.34, the balance has clearly shifted. I have no idea why this happened and am not interested, but the probable conclusion from the tests is that s6.32 is currently the best we can do.
Yes, but the trend changes. So I don't think this will be the best network.
@michito744 Will it be a good idea to estimate the leading scores in katago's official training and rating games and when it is big there might be a blunder and training these situations when the blunder happens?
like the Resigned games, estimate again to see how much B/W leads and if it's over 30 or so, we can analyse what happened