60b made bad moves sometimes when there were many better choise

[ceremony08]

is it a bug?

http://h5.foxwq.com/txwqshare/index.html?chessid=1635777082030051618&title=%E3%80%90%E7%BD%91%E6%A3%8B%E6%A3%8B%E8%B0%B1%E3%80%91&desc=mst007%5B7%E7%BA%A7%5D%20VS%20%E6%B2%89%E6%80%9D%E6%9C%BA%E5%99%A8%E7%8C%AB%5B9%E6%AE%B5%5D&catelog=2&statisticsShareType=4&clientType=13

katago made many mistake after the 140 move

[michito744]

I think it would be easier to share information if you provide the sgf.

[ceremony08]

I think it would be easier to share information if you provide the sgf.

yes，i will provide sgf later

[kiatchun]

My kataGo TensorRT was playing bad moves at the start by playing at the edge of the board the whole time. It was battling the cuda version. Both were using 60b net. Could the bad moves be because of 60b or TensorRT.

[ceremony08]

MyGames.zip

60b,tensorRT,card RTX3090, windows10

komi 0, 50000visit limit,pda3.0

[michito744]

KataGo is W, isn't it?

[hope366]

I analyzed it using a weight of 60b. Certainly, it seems that katago often selects a move with a low winning percentage after the 140th move. I noticed another strange point. Looking at the screenshot below, at this point katago shows that Black is leading by about 30 points, but after this, Black's lead is 10 points or less no matter which move candidate Black puts a stone on. It will be displayed.

[michito744]

There is a possibility that komi were set during the game.

When analyzed with komi 7.5 (in Chinese rules), there are some unnatural moves, but only moves that are acceptable within the margin of error in winning rate.

If there is a problem, it is on the KataGo itself, where the points difference causes extreme searching variations.

[ceremony08]

KataGo is W, isn't it?

YES, black is a very strong player

[ceremony08]

There is a possibility that komi were set during the game.

When analyzed with komi 7.5 (in Chinese rules), there are some unnatural moves, but only moves that are acceptable within the margin of error in winning rate.

If there is a problem, it is on the KataGo itself, where the points difference causes extreme searching variations.

komi value is 0

[michito744]

@ceremony08 I know that you intended 0 komi, but I think the natural supposition. In my environment, it works perfectly fine, and when I analyze this game record with 0 komi, it makes normal moves.

You might want to report your operating environment as well, just in case.

[ceremony08]

I analyzed it using a weight of 60b. Certainly, it seems that katago often selects a move with a low winning percentage after the 140th move. I noticed another strange point. Looking at the screenshot below, at this point katago shows that Black is leading by about 30 points, but after this, Black's lead is 10 points or less no matter which move candidate Black puts a stone on. It will be displayed.

@ceremony08 I know that you intended 0 komi, but I think the natural supposition. In my environment, it works perfectly fine, and when I analyze this game record with 0 komi, it makes normal moves.

You might want to report your operating environment as well, just in case.

YES, i am busy these days, and i will do some test later

60b,tensorRT,card RTX3090, windows10 komi 0, 50000visit limit,pda3.0

[lightvector]

Are you overclocking your GPU or have you noticed any hardware issues? Once before I've seen a case where someone's hardware was configured such that the GPU was overloaded, but the overloading did not cause any outright crashes or failures, instead KataGo just made bad moves and returned anomalous scores sometimes.