Adversarial policies in Go

Per-iteration adversarial training progress

Here we show games from each iteration of our iterated adversarial training procedure. At each iteration, the victim vn is generated by fine-tuning the previous victim vn-1 against the previous adversary an-1, and the adversary an is generated by fine-tuning the previous adversary an-1 against the victim vn. The initial victim v0 is KataGo network checkpoint b40c256-s11840935168-d2898845681, the network we attacked with our original cyclic adversary (base-adversary). The initial adversary a0 is the original cyclic adversary.

a0 vs. v0

Against victim v0, a0 achieves win rates of 100% at 16 victim visits, 99% at 256 visits, and 99% at 4096 visits.

Victim
Rank
-
Caps
4
Time
--:--
Adversary
Rank
-
Caps
49
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -108.2

Victim: v0, 16 visits

Adversary: a0

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue108.5304
wbTrue108.5272
wbTrue110.5280
wbTrue114.5272
wbTrue116.5289
wbTrue124.5305
wbTrue136.5351
bwTrue117.5311
bwTrue125.5305
bwTrue127.5308

a0 vs. v1

Against victim v1, a0 achieves win rates of 13% at 16 victim visits, 1% at 256 visits, and 1% at 4096 visits.

Adversary
Rank
-
Caps
56
Time
--:--
Victim
Rank
-
Caps
47
Time
--:--
Comments
Black passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -53.0

Victim: v1, 16 visits

Adversary: a0

Victim ColorWin colorAdversary WinScore differenceGame length Download
bwTrue53.5335
bwTrue105.5310
wwFalse-257.5382
wwFalse-231.5386
wwFalse-215.5407
bbFalse-304.5446
bbFalse-276.5389
bbFalse-248.5429
bbFalse-248.5379
bbFalse-220.5411

a1 vs. v1

Against victim v1, a1 achieves win rates of 99% at 16 victim visits, 86% at 256 visits, and 69% at 4096 visits.

Victim
Rank
-
Caps
25
Time
--:--
Adversary
Rank
-
Caps
39
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -22.7

Victim: v1, 16 visits

Adversary: a1

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue22.5306
wbTrue104.5361
wbTrue108.5298
wbTrue134.5350
wbTrue226.5441
wbTrue278.5459
bwTrue33.5319
bwTrue149.5440
bwTrue175.5420
bwTrue277.5434

a1 vs. v2

Against victim v2, a1 achieves win rates of 34% at 16 victim visits, 4% at 256 visits, and 0% at 4096 visits.

Victim
Rank
-
Caps
11
Time
--:--
Adversary
Rank
-
Caps
43
Time
--:--
Comments
Black passed.

adversary predicted win prob: 1.00 loss: 0.00, predicted score: 68.5

Victim: v2, 16 visits

Adversary: a1

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue68.5273
wbTrue184.5433
bwTrue43.5283
bwTrue67.5301
wwFalse-199.5360
wwFalse-169.5314
wwFalse-133.5356
bbFalse-150.5327
bbFalse-122.5261
bbFalse-14.5291

a2 vs. v2

Against victim v2, a2 achieves win rates of 99% at 16 victim visits, 95% at 256 visits, and 56% at 4096 visits.

Victim
Rank
-
Caps
15
Time
--:--
Adversary
Rank
-
Caps
103
Time
--:--
Comments
Black passed.

adversary predicted win prob: 1.00 loss: 0.00, predicted score: 210.6

Victim: v2, 16 visits

Adversary: a2

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue210.5381
wbTrue286.5453
wbTrue290.5413
wbTrue354.5447
wbTrue354.5433
bwTrue137.5402
bwTrue171.5332
bwTrue243.5468
bwTrue313.5383
bwTrue367.5458

a2 vs. v3

Against victim v3, a2 achieves win rates of 13% at 16 victim visits, 4% at 256 visits, and 0% at 4096 visits.

Adversary
Rank
-
Caps
75
Time
--:--
Victim
Rank
-
Caps
10
Time
--:--
Comments
Black passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -176.5

Victim: v3, 16 visits

Adversary: a2

Victim ColorWin colorAdversary WinScore differenceGame length Download
bwTrue175.5349
wwFalse-197.5404
wwFalse-192.5395
wwFalse-179.5353
wwFalse-171.5400
wwFalse-57.5300
bbFalse-256.5437
bbFalse-221.5386
bbFalse-212.5417
bbFalse-198.5395

a3 vs. v3

Against victim v3, a3 achieves win rates of 88% at 16 victim visits, 67% at 256 visits, and 33% at 4096 visits.

Victim
Rank
-
Caps
11
Time
--:--
Adversary
Rank
-
Caps
42
Time
--:--
Comments
Black passed.

adversary predicted win prob: 1.00 loss: 0.00, predicted score: 52.3

Victim: v3, 16 visits

Adversary: a3

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue52.5307
wbTrue62.5298
wbTrue80.5324
bwTrue23.5279
bwTrue59.5291
bwTrue71.5325
bwTrue85.5297
bwTrue95.5289
wwFalse-159.5308
wwFalse-3.5354

a3 vs. v4

Against victim v4, a3 achieves win rates of 18% at 16 victim visits, 0% at 256 visits, and 0% at 4096 visits.

Victim
Rank
-
Caps
39
Time
--:--
Adversary
Rank
-
Caps
45
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -10.0

Victim: v4, 16 visits

Adversary: a3

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue10.5340
wbTrue24.5312
wwFalse-232.5396
wwFalse-163.5290
wwFalse-156.5341
wwFalse-155.5332
bbFalse-194.5335
bbFalse-182.5321
bbFalse-134.5291
bbFalse-102.5247

a4 vs. v4

Against victim v4, a4 achieves win rates of 94% at 16 victim visits, 54% at 256 visits, and 25% at 4096 visits.

Victim
Rank
-
Caps
27
Time
--:--
Adversary
Rank
-
Caps
57
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -14.3

Victim: v4, 16 visits

Adversary: a4

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue14.5344
wbTrue20.5330
wbTrue24.5374
wbTrue50.5322
bwTrue23.5279
bwTrue23.5287
bwTrue29.5331
bwTrue39.5377
bbFalse-77.5379
bbFalse-26.5335

a4 vs. v5

Against victim v5, a4 achieves win rates of 9% at 16 victim visits, 0% at 256 visits, and 0% at 4096 visits.

Adversary
Rank
-
Caps
65
Time
--:--
Victim
Rank
-
Caps
36
Time
--:--
Comments
Black passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -23.0

Victim: v5, 16 visits

Adversary: a4

Victim ColorWin colorAdversary WinScore differenceGame length Download
bwTrue23.5345
wwFalse-140.5364
wwFalse-129.5312
wwFalse-123.5296
wwFalse-121.5350
wwFalse-106.5315
wwFalse-83.5266
bbFalse-118.5331
bbFalse-110.5351
bbFalse-108.5295

a5 vs. v5

Against victim v5, a5 achieves win rates of 51% at 16 victim visits, 3% at 256 visits, and 0% at 4096 visits.

Victim
Rank
-
Caps
25
Time
--:--
Adversary
Rank
-
Caps
61
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -12.1

Victim: v5, 16 visits

Adversary: a5

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue12.5326
wbTrue16.5322
wbTrue30.5328
bwTrue3.5291
bwTrue21.5342
bwTrue47.5321
wwFalse-149.5339
bbFalse-162.5307
bbFalse-152.5313
bbFalse-150.5365

a5 vs. v6

Against victim v6, a5 achieves win rates of 10% at 16 victim visits, 1% at 256 visits, and 0% at 4096 visits.

Victim
Rank
-
Caps
13
Time
--:--
Adversary
Rank
-
Caps
56
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -6.6

Victim: v6, 16 visits

Adversary: a5

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue8.5286
bwTrue7.5315
wwFalse-170.5351
wwFalse-167.5372
wwFalse-129.5330
wwFalse-119.5356
bbFalse-198.5377
bbFalse-170.5393
bbFalse-162.5381
bbFalse-116.5408

a6 vs. v6

Against victim v6, a6 achieves win rates of 38% at 16 victim visits, 9% at 256 visits, and 2% at 4096 visits.

Adversary
Rank
-
Caps
55
Time
--:--
Victim
Rank
-
Caps
18
Time
--:--
Comments
Black passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -17.4

Victim: v6, 16 visits

Adversary: a6

Victim ColorWin colorAdversary WinScore differenceGame length Download
bwTrue17.5307
bwTrue37.5329
bwTrue41.5383
wwFalse-266.5356
wwFalse-161.5308
wwFalse-155.5320
wwFalse-133.5306
wwFalse-77.5302
bbFalse-200.5355
bbFalse-164.5389

a6 vs. v7

Against victim v7, a6 achieves win rates of 26% at 16 victim visits, 1% at 256 visits, and 0% at 4096 visits.

Victim
Rank
-
Caps
50
Time
--:--
Adversary
Rank
-
Caps
72
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -10.6

Victim: v7, 16 visits

Adversary: a6

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue10.5378
wbTrue68.5372
wbTrue72.5334
bwTrue59.5357
wwFalse-367.5386
wwFalse-261.5386
wwFalse-137.5320
bbFalse-166.5333
bbFalse-164.5363
bbFalse-152.5265

a7 vs. v7

Against victim v7, a7 achieves win rates of 94% at 16 victim visits, 47% at 256 visits, and 32% at 4096 visits.

Victim
Rank
-
Caps
19
Time
--:--
Adversary
Rank
-
Caps
64
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -38.6

Victim: v7, 16 visits

Adversary: a7

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue38.5346
wbTrue38.5352
wbTrue40.5366
wbTrue42.5356
wbTrue50.5376
wbTrue64.5346
wbTrue68.5342
wbTrue70.5350
bwTrue35.5359
bwTrue67.5347

a7 vs. v8

Against victim v8, a7 achieves win rates of 16% at 16 victim visits, 0% at 256 visits, and 0% at 4096 visits.

Adversary
Rank
-
Caps
66
Time
--:--
Victim
Rank
-
Caps
19
Time
--:--
Comments
Black passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -34.9

Victim: v8, 16 visits

Adversary: a7

Victim ColorWin colorAdversary WinScore differenceGame length Download
bwTrue35.5351
wwFalse-367.5436
wwFalse-273.5384
wwFalse-221.5280
wwFalse-214.5392
wwFalse-213.5352
bbFalse-354.5443
bbFalse-279.5407
bbFalse-270.5421
bbFalse-168.5335

a8 vs. v8

Against victim v8, a8 achieves win rates of 24% at 16 victim visits, 0% at 256 visits, and 0% at 4096 visits.

Adversary
Rank
-
Caps
66
Time
--:--
Victim
Rank
-
Caps
52
Time
--:--
Comments
Black passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -21.0

Victim: v8, 16 visits

Adversary: a8

Victim ColorWin colorAdversary WinScore differenceGame length Download
bwTrue21.5371
bwTrue39.5347
wwFalse-333.5373
wwFalse-310.5401
wwFalse-292.5381
wwFalse-207.5370
wwFalse-49.5407
bbFalse-274.5411
bbFalse-228.5343
bbFalse-222.5429

a8 vs. v9

Against victim v9, a8 achieves win rates of 6% at 16 victim visits, 2% at 256 visits, and 0% at 4096 visits.

Victim
Rank
-
Caps
221
Time
--:--
Adversary
Rank
-
Caps
50
Time
--:--
Comments
White passed.

victim predicted win prob: 1.00 loss: 0.00, predicted score: 366.5

Victim: v9, 16 visits

Adversary: a8

Victim ColorWin colorAdversary WinScore differenceGame length Download
wwFalse-367.5486
wwFalse-367.5360
wwFalse-325.5448
wwFalse-323.5426
wwFalse-307.5418
wwFalse-289.5422
bbFalse-354.5351
bbFalse-278.5391
bbFalse-241.5411
bbFalse-206.5385

a9 vs. v9

Against victim v9, a9 achieves win rates of 99% at 16 victim visits, 94% at 256 visits, and 59% at 4096 visits.

Victim
Rank
-
Caps
10
Time
--:--
Adversary
Rank
-
Caps
62
Time
--:--
Comments
White passed.

victim predicted win prob: 0.00 loss: 1.00, predicted score: -60.1

Victim: v9, 16 visits

Adversary: a9

Victim ColorWin colorAdversary WinScore differenceGame length Download
wbTrue60.5294
wbTrue64.5310
wbTrue88.5324
wbTrue102.5326
bwTrue57.5307
bwTrue57.5337
bwTrue59.5287
bwTrue123.5315
bwTrue137.5347
bwTrue171.5371

Citation Info

@inproceedings{wang2023adversarial,
  title={Adversarial Policies Beat Superhuman Go {AI}s},
  author={Wang, Tony T. and Gleave, Adam and Tseng, Tom and Pelrine, Kellin and Belrose, Nora and Miller, Joseph and Dennis, Michael D and Duan, Yawen and Pogrebniak, Viktor and Levine, Sergey and Russell, Stuart},
  booktitle={International Conference on Machine Learning},
  year={2023},
  eprint={2211.00241},
  archivePrefix={arXiv}
}
@misc{tseng2024ais,
  title={Can Go {AI}s be adversarially robust?},
  author={Tseng, Tom and McLean, Euan and Pelrine, Kellin and Wang, Tony T. and Gleave, Adam},
  year={2024},
  eprint={2406.12843},
  archivePrefix={arXiv}
}