[Computer-go] guess AlphaGo Zero's CGOS rating

Hiroshi Yamashita yss at bd.mbn.or.jp
Thu Jan 25 23:45:08 PST 2018


Hi,

I guessed AlphaGo Zero's CGOS rating.
I picked up ratings from "Mastering the Game of Go without Human Knowledge"
 Figure 3 and Figure 6.

                           GoRatings
Raw network                3030  (40 blocks)  (Policy)
AlphaGo Zero               5200  (40 blocks)  1 machine, 4 TPU
AlphaGo Master             4800               1 machine, 4 TPU, Rollout
AlphaGo Zero               4400  (20 blocks)  1 machine, 4 TPU
AlphaGo Lee                3700               many machines,  48 TPU  2016/03
AlphaGo Fan                3150               many machines, 176 GPU  2015/10
CrazyStone                 1900
Pachi                      1300
GNU Go                      400

I think over AlphaGo Lee, ratings calculated from selfplay. So I estimate
 rating progress is half. For example, Zero(5200) is +1500 stronger than
 Lee(3700). But I estimate Zero is +750 stronger ,Zero(4450).

                  BayesElo
                   CGOS    GoRatings
AlphaGo Zero       5422?   4450? (40 blocks)  1 machine, 4 TPU
AlphaGo Master     5231?   4250?              1 machine, 4 TPU, Rollout
AlphaGo Zero       5022?   4050? (20 blocks)  1 machine, 4 TPU
AlphaGo Lee        4672?   3700               many machines, 48 TPU   2016/03
Park Junghwan      4592?   3620  strongest human
Ke Jie             4590?   3618
Yuta Iyama         4546?   3574
Lee Sedol          4514?   3542
Zen-15.7-4c1g      4269    3297? CGOS strongest
Cho Chikun         4187?   3215
AlphaGo Fan        4122?   3150               many machines, 176 GPU  2015/10
tianrang-galaxy    4062    3090?
Raw network        4002?   3030  (40 blocks)  (Policy)
Fan Hui            3978?   3006
AQ-2.1.1-4t1g      3944    2972?
AlphaGo(Fan)       3862?   2890               1 machine(48CPU, 8GPU) 2015/10
Rn.4.16-4c         3525    2553?
Aya792p2v2cn50_12t 3313    2341? KGS 7d (Blitz)
AlphaGo RL         2729?   1757? Reinforcement learning from AlphaGo SL
Aya790e_510_ro_1k  2580    1608? KGS 3d
AlphaGo SL         2489?   1517               (Policy) Filter 192
pachi11_12t_100k   2429    1457? pachi, 100k playout/move
Gnugo-3.7.10-a1    1800     828? KGS 5k

Note 1. CGOS <-> GoRatings is based on AlphaGo RL beats Pachi 100k with 85%(+300).
Note 2. CGOS = GoRatings + 972
Note 3. "?" is guess.
Note 4. AlphaGo Lee(3700) may be calculated from 4-1 result against Lee Sedol(3542).
Note 5. Human ratings are from https://www.goratings.org/en/
Note 6. CGOS rating is BayesElo. http://www.yss-aya.com/cgos/19x19/bayes.html

Thanks,
Hiroshi Yamashita



More information about the Computer-go mailing list