Engine Tournament

[q30]

KataGo v. 1.14.1 - 1.14.0: 25 - 21 (details).

[q30]

New KataGo weight files of "light heavyweight category" are weaker than only "middleweight" ones (details).

[q30]

The current rate of KataGo "weight categories":
"bantamweight" <|= 2 ^ 23 B (< 12 MiB) - g170e-b10c128-s1141046784-d204142634.bin (6)
"featherweight" 2 ^ 24 B (12 - 24 MiB) - I haven't
"lightweight" 2 ^ 25 B (24 - 48 MiB) - g170e-b15c192-s1672170752-d466197061.bin (5)
"welterweight" 2 ^ 26 B (48 - 96 MiB) - g170e|kata1-b20c256x2-s5303129600-d1228401921.bin (4)
"middleweight" 2 ^ 27 B (96 - 192 MiB) - kata1-b18c384nbt-s9131461376-d4087399203.bin (1)
"light heavyweight" 2 ^ 28 B (192 - 384 MiB) - b28c512nbt-s5668008960-d4210144556.bin (2)
"heavyweight" 2 ^ 29 B (384 - 768 MiB) - kata1-b60c320-s6782286336-d3070935549.bin (3)
"super heavyweight" >|= 2 ^ 30 B (> 768 MiB) - I haven't

[xela]

New KataGo weight files of "light heavyweight category" are weaker than only "middleweight" ones (details).

It looks like you're using very fast time limits. In slower games, the larger networks get more value from the extra time and will become relatively stronger. Compare the 1-minute and 5-minute rankings at https://lifein19x19.com/viewtopic.php?p=248817#p248817 and see how kata_20b overtakes kata_15b by a large margin given more time (and similar but less drastically for LZ188, 40 blocks versus LZ157, 15 blocks).

Heavyweights are slower but more powerful :-)

[Mike Novack]

And with computers, not absolute time but time and hardware.

We really are going to have to come to some agreement about those. What will be considered a "standard machine" and what "standard time controls.

And what sort of time controls? Allowing X per move is simple but I believe we will eventually want AI that can be making "time management" decisions << situation not critical; try to conserve time vs critical, use some of the saved time >>

[q30]

Ray v. 11.1.0 - the strongest v. (from 02.09.19): 7 - 13 (details).

[q30]

It looks like you're using very fast time limits. In slower games, the larger networks get more value from the extra time and will become relatively stronger. Compare the 1-minute and 5-minute rankings at https://lifein19x19.com/viewtopic.php?p=248817#p248817 and see how kata_20b overtakes kata_15b by a large margin given more time (and similar but less drastically for LZ188, 40 blocks versus LZ157, 15 blocks).

Heavyweights are slower but more powerful :-)

I'm testing with 2 minutes per move. You can find number of visits for different weights network files on this and this pages.

Of course You are right. But the goal is to test with equal time and resources to each engine/weight. If You can test, how many time per move for this amount of visits will spend modern PC with 8-core CPU + modern video card GPU cores number, I will only welcome the release of these results data here.
I think, that ~10 seconds per move is that a usual end user uses...

[q30]

And with computers, not absolute time but time and hardware.

We really are going to have to come to some agreement about those. What will be considered a "standard machine" and what "standard time controls.

And what sort of time controls? Allowing X per move is simple but I believe we will eventually want AI that can be making "time management" decisions << situation not critical; try to conserve time vs critical, use some of the saved time >>

Of course You are right too. What I think about "standard" for usual end user (that prefers the computer to spend time thinking evenly), I had wrote in message above...

[q30]

KataGo v.1.15.1-v.1.14.1: 10-14 (details).

[q30]

The KataGo "weight categories" new rate (details):
"bantamweight" <|= 2 ^ 23 B (< 12 MiB) - g170e-b10c128-s1141046784-d204142634.bin (6)
"featherweight" 2 ^ 24 B (12 - 24 MiB) - I haven't
"lightweight" 2 ^ 25 B (24 - 48 MiB) - g170e-b15c192-s1672170752-d466197061.bin (5)
"welterweight" 2 ^ 26 B (48 - 96 MiB) - g170e|kata1-b20c256x2-s5303129600-d1228401921.bin (4)
"middleweight" 2 ^ 27 B (96 - 192 MiB) - kata1-b18c384nbt-s9131461376-d4087399203.bin (2)
"light heavyweight" 2 ^ 28 B (192 - 384 MiB)- kata1-b28c512nbt-s7168446720-d4316919285.bin (1)
"heavyweight" 2 ^ 29 B (384 - 768 MiB)- kata1-b60c320-s6782286336-d3070935549.bin (3)
"super heavyweight" >|= 2 ^ 30 B (> 768 MiB)- I haven't

[q30]

The higher probability of "human like" play in KataGo - the weaker play (details).

[q30]

Ray got the strogest level of its play, that was in 2021 year (details).