Sunteți pe pagina 1din 37

ALPHAGO ZERO DETHRONED?

Comparison of AlphaGo Zero (both 20 & 40 residual blocks sized) with its Deep
Reinforcement Learning Clones
Based on Matching the 40 Samples of AlphaGo Zero’s Moves

1. INTRODUCTION

There is an urban legend in a Go world about famous AlphaGo algorithm. Almost


everyone agree – but no one knows for sure – that breakthrough zero human knowledge versions
of AlphaGo (AlphaGo Zero & AlphaZero) are the strongest Go game AI's ever played.
We can call it some kind of primacy effect, when the First One (who has beaten the top
human pro's in our case) stucks in collective mind as the toughest one for a long time.
Same thing that happened with Bruce Lee, who brought the Eastern martial arts to the
Western mass culture. Exactly like AlphaGo algorithm brought the AI way of playing to the Go
game mainstream but in backward geographical direction.
The main question is does the reputation correspond to the real world?
Could Bruce Lee beat the strongest contemporary rivals in early 1970-th, such as world
boxing champion Joe Frazier? Could AlphaGo Zero beat FineArt in late 2017? In other words,
is the Google DeepMind’s bot really the all-time best?
Everyone thinks, “Hell yeah!”1
However, no one still doesn't know, what if... None of these dream fights never took a
place. Bruce wasn't a professional fighter, and DeepMind's team didn't have enough motivation
to take a risk and compete with another AI after they left the human intelligence behind and got
their portion of super hype.
That's why the only thing left for us is to make a reasonable comparison. Just take a few
samples of game material considerable by many as a nonpareil benchmark, and take a look at
how close the post AlphaGo open source AI's are able to match it.
Due to the absence of AlphaZero games, which DeepMind's digital Dark Tower still
keeping imprisoned, I use its predecessor AlphaGo Zero as a personification of perfect play
inside the 19x19 intersections battlefield.
The main hypothesis is simple: the urban legend about unparalleled AlphaGo game
strength is correct. Facebook AI researchers (the ELF OpenGo team) agree with AGZ toughness,
although they have no access to direct comparison as they had it with Leela Zero 15B model.
Here is their assessment of AlphaGo as a priority benchmark: “After running our AlphaZero-
style training software on 2,000 GPUs for 9 days, our 20-block model has achieved superhuman
performance that is arguably comparable to the 20-block models described in Silver et al. (2017)
and Silver et al. (2018)”2.
Mark it one more time: in theory, the newest ELF v2 version considered by its developers
as arguably comparable only to 20B AlphaGo Zero & AlphaZero. No hint about AGZ 40B.
That's why the more matches to AlphaGo Zero game samples among the recent AI's will
be found the stronger we should consider them. It is logical to expect more matches from newer
versions of bots because of a) today they are stronger than yesterday, and b) they are finding the
best moves faster nowadays.

1
E.g:“Alphazero, developed by Google, is the strongest known AI at go and several other board games, and its design serves as the
template for the other top AI such as leelazero.” Hu (2019)
2
Look at ELF OpenGo: An Analysis and Open Reimplementation of AlphaZero, Feb 2019.
Hypothesis limits are simple as well. Unexpected upset with more matches from older
and weaker AI is able to ruin the legend and push AlphaGo Zero away from its throne of hype.
No internal corporative rating system can be used as an executioner’s axe instead. As a
Leela Zero's creator Gian-Carlo Pascutto says, closed selfplay based ranking numbers are not
trustworthy.
Nevertheless, I do not expect to see AlphaGo dethroned as result of my small comparative
analysis. At least in the name of the hype power.

2. SETTING THE SCENE

Let's set the comparison.


First of all, let’s pick the AlphaGo Zero moves samples from its latest games, both for 20
& 40 residual blocks sizes. According to aforementioned ELF OpenGo team assessment, we
should expect more matches for 20B AGZ model as a weaker one.
The guitanj’s AutoPK Elo-playouts rating calculations confirms this assumption. The AI
rankings he got for Leela Zero, ELF OpenGo and MiniGo, shows 3500 to 4000 Elo points for
15 & 20 blocks Networks within 1600 – 3200 playouts range, and 3800 to 4200 Elo points for
LZ 40 residual blocks iterations within the same number of readings.
I’ve selected five moves from each of eight picked games (50/50 to both sizes Networks).
All of them are beyond move 30. The selfplay randomization argument –m 30 stops there3 and
AlphaGo Zero training games becomes as close to its full strength play as it possible, also taking
into account adding the Dirichlet noise to AGZ’s MCTS exploration process during selfplay4.
Here is my choice for AlphaGo Zero 20B:

Game 17: moves 34 – 38 Game 18: moves 33, 36, 46, 53, 79

3
“For the first 30 moves of each [selfplay] game, the temperature is set to τ→ 1; this selects moves proportionally to their visit count
in MCTS, and ensures a diverse set of positions are encountered. For the remainder of the game, an infinitesimal temperature is used,
τ→ 0… i.e. we deterministically select the move with maximum visit count, to give the strongest possible play.” Silver et al. (2017)
4
“Additional exploration is achieved by adding Dirichlet noise to the prior probabilities in the root node s0… this noise ensures that
all moves may be tried, but the search may still overrule bad moves.” Silver et al. (2017)
Game 19: moves 43, 44, 45, 70, 72 Game 20: moves 36, 38, 39, 41, 43

Here are the picked moves for AlphaGo Zero 40B:

Game 17: moves 33, 34, 42, 44, 56 Game 18: moves 35, 50, 56, 79, 106
Game 19: moves 32, 39, 41, 62, 71 Game 20: moves 46, 50, 59, 105, 106

Secondly, let’s fill the list of top open source Zero Networks by some milestone releases
for the last year. One non-Zero, SL Network added as a checkpoint to estimate how far AI’s RL
generation is ahead of the first wave of Go playing Deep Learning bots5.

RELEASE ENGINE NETWORK


2017-11 Leela 0.11 DCNN – checkpoint
2018-04 LZ 0.17 Leela Zero #116, 10B
2018-04 LZ 0.17 Leela Zero #128, 15B
2018-05 LZ 0.17 ELF OpenGo v0, 20B
2018-05 LZ 0.15* PhoenixGo v1, 20B
2018-07 LZ 0.17 Leela Zero #157, 15B
2018-07 LZ 0.17 ELF OpenGo v1, 20B
2018-09 LZ 0.17 Leela Zero #173, 20B
2018-09 LZ 0.17 Leela Zero #174, 40B
2018-11 LZ 0.17 Leela Zero #188, 40B
2018-12 LZ 0.17 MiniGo v14 #979, 20B
2019-01 LZ 0.17 MiniGo v15 #990, 20B
2019-02 LZ 0.17 ELF OpenGo v2, 20B
2019-02 LZ 0.17 Leela Zero #204, 15B trained on 40B
2019-02 LZ 0.17 MiniGo v16 #1144, 40B
2019-03 LZ 0.16* MiniGo v17 #961, 20B
2019-05 LZ 0.17 Leela Zero #224 + ELF v2, 15B trained on 40B
2019-05 LZ 0.17 Leela Zero #226, 40B

5
“Leela contains an AI technique modeled after the human visual system and brain, a deep stack of artificial neurons processing visual
input with each layer combining the previous ones into higher level features (a so called DCNN, deep convolutional neural network).
This "neural network" has been trained with more than 32 million positions from high level go games and taught to predict which
moves a pro player would most likely consider. In 19x19 games the engine will query this deep "neural network" during the search to
focus on the most critical variations. The result is a substantial raise in playing strength (about 6 stones), and a more human-like
playstyle, while still allowing the engine to innovate of its own.” Sjeng.Org
Engines marked with * have been modified because there was no another way to adopt
those AI’s weights to the current original Leela Zero engine.
In addition, we should remember that both ELF OpenGo and PhoenixGo Networks for
2018 were far stronger than synchronous Leela Zero & MiniGo iterations.
In order to reflect this fact and maximize the result’s informativeness it makes sense to
divide the AI list into several categories, in accordance with the algorithm’s game strength.
# CATEGORY MACHINE LEARNING ALGORITHMS
1 State-of-the-art ELF OpenGo v2, 20B
MiniGo v17 #961, 20B
Leela Zero #224 + ELF v2, 15B trained on 40B
Leela Zero #226, 40B
2 AI’s from yesterday ELF OpenGo v0 & v1, 20B
PhoenixGo v1, 20B
Leela Zero #174 & 188, 40B
Leela Zero #204, 15B trained on 40B
MiniGo v15 #990, 20B
MiniGo v16 #1144, 40B
3 AI’s from day before yesterday Leela Zero #157, 15B
Leela Zero #173, 20B
MiniGo v14 #979, 20B
4 Blast from the past Leela Zero #116, 10B
Leela Zero #128, 15B
Finally yet importantly, let’s define the comparison's options.
AlphaGo Zero has been using 1.6k simulations per move both for selfplay and evaluation
games6. According to moves analyzing experience with GUI Lizzie, AI normally spent most
part of these MCTS visits number for exploring the #1 move (the blue one). This proportion is
able to reach up to 100% of all simulations for the first line move.
No AlphaGo Zero moves known for such a games with the higher number of visits. AI
analysis with Lizzie shows: higher calculations always drives to completely different sequences
of moves. That’s why any attempt to use significantly bigger number of visits is unreasonable.
From author’s point of view, the best way for this comparison is to set 1.6k simulations
specifically for #1 move (the blue one). If the blue move is behind in visits count, MCTS stops
by picking the first one visited 1.6k times, which still is in accordance with AlphaGo Zero
paper’s idea.
It might be a bit higher number of readings than some of AGZ’s moves had, but looks
like pretty clear and appropriate standard.
The evaluation system of AlphaGo Zero moves matching is simple: *** - the blue move
with 1.6k visits, or a non-blue one which turns 1.6k visits first, matches AGZ; ** - no top move
with 1.6k visits matches AGZ, but AI considers Alpha’s move as the most promising one (light
blue); * - top move does not match AGZ, but Alpha’s move is one of the non-blue candidate
ones; “-” – no candidate move matches AlphaGo Zero.
Amount of stars is equal to the points it gives to each AI.
6
For selfplay: “The best current player αө*, as selected by the evaluator, is used to generate data. In each iteration, αө* plays 25,000
games of self-play, using 1,600 simulations of MCTS to select each move (this requires approximately 0.4s per search)”. For
evaluation games: “To ensure we always generate the best quality data, we evaluate each new neural network checkpoint against the
current best network fө before using it for data generation… Each evaluation consists of 400 games, using an MCTS with 1,600
simulations to select each move”. Silver et al. (2017)
3. COMPARATIVE ANALYSIS ITSELF

Sample 1: AlphaGo Zero 20B, game 17

AI’S 34.n10 35.o11 36.n13 37.r8 38.m5 PTS


Leela 0.11 – checkpoint * * - - * 3
Leela Zero #116, 10B - * - * *** 5
Leela Zero #128, 15B - *** - * *** 7
ELF OpenGo v0, 20B * *** - * * 6
PhoenixGo v1, 20B * *** - * *** 8
Leela Zero #157, 15B ** ** - * *** 8
ELF OpenGo v1, 20B * *** - * *** 8
Leela Zero #173, 20B * ** - * * 5
Leela Zero #174, 40B * *** - * *** 8
Leela Zero #188, 40B * *** - * *** 8
MiniGo v14 #979, 20B - * - * *** 5
MiniGo v15 #990, 20B - *** - *** * 7
ELF OpenGo v2, 20B * *** - ** *** 9
Leela Zero #204, 15B/40B * *** - * *** 8
MiniGo v16 #1144, 40B - *** - * * 5
MiniGo v17 #961, 20B - *** - *** * 7
Leela Zero #224/ELF v2, 15B/40B * *** - * *** 8
Leela Zero #226, 40B ** *** - * *** 9

34.n10: the move no one picked as a top one. Only two Leela’s iterations consider move
34 as the most promising one. The first is a 15 blocks “long play” #157:
The second is #226, the record-breaking model that wiped out 53 newer Networks:

35.p11: the move only two Zero AI’s cannot see as a blue or a light blue one. The last 10
blocks iteration of Leela Zero #116 doesn’t like this keima very much:
As well as the strongest MiniGo v14 Network, #979-two-lions:

36.n13: the move invisible for everyone. The most genius or the noisiest one? Even the
All-seeing Eye of PhoenixGo v1 was not able to detect it:
37.r8: the move which been picked only by two last MiniGo 20 blocks models, v15 and
v17, despite all MiniGo models in the list were blind during the hunt for move 34. MiniGo v15:

MiniGo v17 can see this knight’s move too:


Sample 2: AlphaGo Zero 20B, game 18

AI’S 33.r2 36.r9 46.o3 53.m6 79.r13 PTS


Leela 0.11 – checkpoint * - * * - 3/6
Leela Zero #116, 10B *** * - *** * 8/13
Leela Zero #128, 15B *** * * *** * 9/16
ELF OpenGo v0, 20B *** * *** *** * 11/17
PhoenixGo v1, 20B *** * *** *** * 11/19
Leela Zero #157, 15B *** * *** *** * 11/19
ELF OpenGo v1, 20B *** * *** *** * 11/19
Leela Zero #173, 20B *** * *** *** * 11/16
Leela Zero #174, 40B *** * *** *** * 11/19
Leela Zero #188, 40B *** * *** *** * 11/19
MiniGo v14 #979, 20B * - *** *** * 8/13
MiniGo v15 #990, 20B *** - *** *** * 10/17
ELF OpenGo v2, 20B *** * *** *** * 11/20
Leela Zero #204, 15B/40B *** * *** *** * 11/19
MiniGo v16 #1144, 40B *** * *** *** * 11/16
MiniGo v17 #961, 20B *** - *** *** * 10/17
Leela Zero #224/ELF v2, 15B/40B *** * *** *** * 11/19
Leela Zero #226, 40B *** * *** *** * 11/20

36.r9: the move every Zero can see except three out of four MiniGo’s models. Looks like
such a blind spots are the family trait of MiniGo project. MiniGo v14 can see nothing here:
As well as MiniGo v15:

Including MiniGo v17, the best one in family for matching the AlphaGo Zero 20B moves:
This eyeless performance correlates with limited moves search proficiency of SL Leela
0.11, living transitional fossil between the old fashion MCTS engines and modern RL Zero bots7.

46.o3: the most logical move? Only the oldest ones among all the chosen AI’s cannot see
it as the #1 move. Leela Zero #116 10B is completely blind on a given simulations:

7
“Much progress towards artificial intelligence has been made using supervised learning systems that are trained to replicate the
decisions of human experts. However, expert data is often expensive, unreliable, or simply unavailable. Even when reliable data is
available it may impose a ceiling on the performance of systems trained in this manner”. Silver et al. (2017)
Leela Zero #128 15B, the benchmark test for ELF OpenGo v0, can see no reason to spend
plenty of readings for this move:

79.r13: the move everyone can see – except the obsolete checkpoint Network – but no
one prefer within the whole one-year range. Whether it initial ELF OpenGo v0:
…or the hybrid bjiyxo Network 9006c708, trained on both LZ #224 and ELF v2 selfplay:

Sample 3: AlphaGo Zero 20B, game 19

AI’S 43.r14 44.n16 45.p14 70.s8 72.q4 PTS


Leela 0.11 – checkpoint * - * * - 3/9
Leela Zero #116, 10B * ** * * * 6/19
Leela Zero #128, 15B * * * * *** 7/23
ELF OpenGo v0, 20B *** - * * * 6/23
PhoenixGo v1, 20B *** * * *** * 9/28
Leela Zero #157, 15B *** - * *** * 8/27
ELF OpenGo v1, 20B *** - * *** * 8/28
Leela Zero #173, 20B *** - * *** *** 10/26
Leela Zero #174, 40B *** - * *** * 8/27
Leela Zero #188, 40B *** - * * * 6/25
MiniGo v14 #979, 20B *** - * * - 5/18
MiniGo v15 #990, 20B *** - * * * 6/23
ELF OpenGo v2, 20B * * * * *** 7/27
Leela Zero #204, 15B/40B *** - * * * 6/25
MiniGo v16 #1144, 40B *** - * * * 6/22
MiniGo v17 #961, 20B *** - * *** *** 10/27
Leela Zero #224/ELF v2, 15B/40B *** - * *** * 8/27
Leela Zero #226, 40B *** - * *** *** 10/30
43.r14: weird fellowship – the newest ELF v2 became the only AI joined to the oldest
two Zeros in the list, plus the checkpoint Network, in their dislike to corner approach 43:

44.n16: the lowest quality move? Silver-haired LZ #116 10B is the only AI that painted
this 4474 enclosure with any shade of blue:
70s8: forgotten 20 blocks Leela’s Network #173 performed unexpectedly strong here:

72.q4: …and further:


Sample 4: AlphaGo Zero 20B, game 20

AI’S 36.h15 38.r4 39.g8 41.r12 43.n4 PTS


Leela 0.11 – checkpoint * * - - *** 5/14
Leela Zero #116, 10B * - *** *** * 8/27
Leela Zero #128, 15B * - *** * * 6/29
ELF OpenGo v0, 20B * * *** * *** 9/32
PhoenixGo v1, 20B * * *** * *** 9/37
Leela Zero #157, 15B * - *** * ** 7/34
ELF OpenGo v1, 20B ** - *** * *** 9/37
Leela Zero #173, 20B * - *** * *** 8/34
Leela Zero #174, 40B ** - *** * * 7/34
Leela Zero #188, 40B * - *** * * 6/31
MiniGo v14 #979, 20B * - *** - *** 7/25
MiniGo v15 #990, 20B * - *** - * 5/28
ELF OpenGo v2, 20B * - *** * *** 8/35
Leela Zero #204, 15B/40B * - *** * * 6/31
MiniGo v16 #1144, 40B *** - *** * *** 10/32
MiniGo v17 #961, 20B * - *** *** *** 10/37
Leela Zero #224/ELF v2, 15B/40B * * *** * *** 9/36
Leela Zero #226, 40B * - *** * *** 8/38

36.h15: MiniGo v16 is the only one AI who was able to catch up this slicky counter-hane.
38.r4: ELF OpenGo v0 paid more attention to this sagari within given simulations than
any other AI from the list:

39.g8: one space jump from Captain Obvious. Even PhoenixGo v1, which likes to explore
larger number of root nodes than any other AI, considers the rest of candidates as pointless ones.
41.r12: another move from the dark side of the Moon dazzles MiniGo. This extension is
out of radars exclusively of MiniGo v14:

…and its successor MiniGo v15:


43.n4: this shoulder hit is the only move that SL Leela 0.11 was able to figure out:

Here we have the subtotal results for matching the AlphaGo Zero 20B moves:

RANK AI’S RELEASE PTS MATCH FOUND


1 Leela Zero #226, 40B 2019-05 38 10/20 17/20
2-4 MiniGo v17 #961, 20B 2019-03 37 11/20 15/20
PhoenixGo v1, 20B 2018-05 37 9/20 19/20
ELF OpenGo v1, 20B 2018-07 37 9/20 17/20
5 Leela Zero #224/ELF v2, 15B/40B 2019-05 36 9/20 18/20
6 ELF OpenGo v2, 20B 2019-02 35 8/20 18/20
7-9 Leela Zero #173, 20B 2018-09 34 8/20 17/20
Leela Zero #174, 40B 2018-09 34 8/20 17/20
Leela Zero #157, 15B 2018-07 34 7/20 17/20
10-11 MiniGo v16 #1144, 40B 2019-02 32 8/20 16/20
ELF OpenGo v0, 20B 2018-05 32 7/20 18/20
12-13 Leela Zero #188, 40B 2018-11 31 7/20 17/20
Leela Zero #204, 15B/40B 2019-02 31 7/20 17/20
14 Leela Zero #128, 15B 2018-04 29 6/20 17/20
15 MiniGo v15 #990, 20B 2019-01 28 7/20 14/20
16 Leela Zero #116, 10B 2018-04 27 5/20 16/20
17 MiniGo v14 #979, 20B 2018-12 25 6/20 13/20
18 Leela 0.11 – checkpoint 2017-11 14 1/20 12/20
SOME OBSERVATIONS ON 20 BLOCKS PART

I) Observations on general compatibility with AlphaGo Zero’s moves

The peak results for Zero AI’s in matching AGZ 20 blocks moves is 55% at the top and
30% at the bottom line. Obviously, not an outstanding performance for Zeros in general.
Herewith, the checkpoint non-Zero Network matched AlphaGo Zero only for minimal 5%.
None of tested Networks wasn’t able to find all 20 moves. Only one of them detects 95%
with the lowest success percentage of 65% for whole AI’s list. Again, the stats might be better.
It would be predictably in light of the main hypothesis (about the unequaled game strength
of DeepMind’s last two Alphas) if just AI’s, which performed superiorly than others, would be
the ones from state-of-the-art category. In other words, ashes to ashes, ELF v2 to AlphaGo Zero.
However, the output we have is not as univocal as it expected to be.

II) Observations on AI’s categories

III) Observations on AI’s families

IV) Conclusions
Sample 5: AlphaGo Zero 40B, game 17

AI’S 33.b3 34.q12 42.f9 49.r9 56.b12 PTS


Leela 0.11 – checkpoint * ** *** * - 7
Leela Zero #116, 10B * * *** * - 6
Leela Zero #128, 15B * ** *** * - 7
ELF OpenGo v0, 20B ** ** *** * * 9
PhoenixGo v1, 20B * * *** * * 7
Leela Zero #157, 15B * ** *** * *** 10
ELF OpenGo v1, 20B ** *** *** * *** 12
Leela Zero #173, 20B * * *** * * 7
Leela Zero #174, 40B * ** *** * * 8
Leela Zero #188, 40B *** * *** * * 9
MiniGo v14 #979, 20B * * *** * * 7
MiniGo v15 #990, 20B ** * *** * *** 10
ELF OpenGo v2, 20B * ** *** * *** 10
Leela Zero #204, 15B/40B *** *** *** * * 11
MiniGo v16 #1144, 40B * * *** * - 6
MiniGo v17 #961, 20B * * *** * - 6
Leela Zero #224/ELF v2, 15B/40B *** ** *** * * 10
Leela Zero #226, 40B * ** *** * * 8

33.b3: only three Leela Zero’s Networks visited this move more than any other candidate.
The first one is #188, breakthrough 40 blocks Net that surpassed #157 in a time parity mode:
…the second one is 15 blocks digital tiger, which has been trained up to LZ #204:

…finally, the third one is a half-blood child of Leela Zero #224 and ELF OpenGo v2:
34.q12: ELF OpenGo v1, probably the strongest open source AI of 2018, became the only
Network from last year that picked this three points approach to the black wall 1-7-9:

49.r9: Leela Zero #157 shows the most popular alternative to AlphaGo Zero’s choice:
56.b12: ELF OpenGo’s last two releases both turned thoughtful enough to figure this
magari out. Here is ELF v1:

…and ELF v2 target list:


Sample 6: AlphaGo Zero 40B, game 18

AI’S 35.d6 50.c2 56.e8 79.p2 106.q13 PTS


Leela 0.11 – checkpoint - ** - - - 2/9
Leela Zero #116, 10B - * * * * 4/10
Leela Zero #128, 15B * *** * * * 7/14
ELF OpenGo v0, 20B *** *** * * * 9/18
PhoenixGo v1, 20B *** *** * * * 9/16
Leela Zero #157, 15B * *** * * * 7/17
ELF OpenGo v1, 20B *** * * * * 7/19
Leela Zero #173, 20B *** *** *** * * 11/18
Leela Zero #174, 40B *** *** * * * 9/17
Leela Zero #188, 40B * *** ** * * 8/17
MiniGo v14 #979, 20B *** * - - * 5/12
MiniGo v15 #990, 20B ** * * * - 5/15
ELF OpenGo v2, 20B ** * * * * 6/16
Leela Zero #204, 15B/40B * *** * *** * 9/20
MiniGo v16 #1144, 40B * * * * * 5/11
MiniGo v17 #961, 20B * * * - * 4/10
Leela Zero #224/ELF v2, 15B/40B * * * * * 5/15
Leela Zero #226, 40B * * *** * * 7/15

35.d6: no AI released since January 1st 2019 doesn’t put this nobi at the top anymore,
unlike their predecessors did. Such as first regular 40 blocks iteration of Leela Zero, #174:
50.c2: similar story with the sagari continuation of this joseki. The sole representative
from nowadays in this case is Leela Zero 15 blocks quantized model #204:

56.e8: the “Leeliest” move among the others? Only Leela’s iterations ## 173 & 226 were
able to push this move to the top of their readings. Here is the last 20 blocks model #173:
…and its modern continuator #226:

79.p2: the total disaster for all the MiniGo models in the list. No match, only two out of
four Networks found this AlphaGo Zero’s 40B defensive move. The blindness of MG v14:
…and surprising helplessness of MG v17 make’em miss the target in Leela 0.11th style:

106.q13: the cut everyone can see except the aged checkpoint Network and MiniGo v15:
Sample 7: AlphaGo Zero 40B, game 19

AI’S 32.n3 39.l16 41.e12 62.n14 71.b18 PTS


Leela 0.11 – checkpoint - - - * *** 4/13
Leela Zero #116, 10B *** * * * * 7/17
Leela Zero #128, 15B *** * * * *** 9/23
ELF OpenGo v0, 20B *** *** ** * * 10/28
PhoenixGo v1, 20B *** *** * * *** 11/27
Leela Zero #157, 15B *** * * * * 7/24
ELF OpenGo v1, 20B *** *** *** * * 11/30
Leela Zero #173, 20B *** *** * * * 9/27
Leela Zero #174, 40B *** *** * * * 9/26
Leela Zero #188, 40B *** *** * * * 9/26
MiniGo v14 #979, 20B * * - * * 4/16
MiniGo v15 #990, 20B *** *** * * * 9/24
ELF OpenGo v2, 20B *** * * * * 7/23
Leela Zero #204, 15B/40B *** * * * *** 9/29
MiniGo v16 #1144, 40B *** *** ** * * 10/21
MiniGo v17 #961, 20B *** * *** * *** 11/21
Leela Zero #224/ELF v2, 15B/40B *** *** ** * *** 12/27
Leela Zero #226, 40B *** *** *** * *** 13/28

32.n3: the hypermodern AI-style attachment tastes utterly bad for an old Leela 0.11 only:
39.l16: LZ #226 is on its sensational rush for matching 4/5 AGZ moves in game 19:

41.e12: once again ELF OpenGo v1 became the only AI, released last year, that picked
exact AlphaGo Zero’s move:
71.b18: v17 is the lone MiniGo version that hit the mark with this atari:

Sample 8: AlphaGo Zero 40B, game 20


AI’S 46.c3 50.r6 59.f2 105.g8 106.d7 PTS
Leela 0.11 – checkpoint * - * - * 3/16
Leela Zero #116, 10B *** - * - * 5/22
Leela Zero #128, 15B *** - * - * 5/28
ELF OpenGo v0, 20B * - *** * *** 8/36
PhoenixGo v1, 20B * * * * *** 7/34
Leela Zero #157, 15B * * * - *** 6/30
ELF OpenGo v1, 20B *** - *** * * 8/38
Leela Zero #173, 20B * - *** - *** 7/34
Leela Zero #174, 40B ** - *** * * 7/33
Leela Zero #188, 40B * - *** * *** 8/34
MiniGo v14 #979, 20B *** - *** - *** 9/25
MiniGo v15 #990, 20B *** - * - *** 7/31
ELF OpenGo v2, 20B * - *** * *** 8/31
Leela Zero #204, 15B/40B ** - *** - *** 8/37
MiniGo v16 #1144, 40B *** - *** - * 7/28
MiniGo v17 #961, 20B *** - *** - *** 9/30
Leela Zero #224/ELF v2, 15B/40B * - ** - *** 6/33
Leela Zero #226, 40B * - *** * * 6/34

46.c3: the last Leela Zero iteration from the list who likes this corner attachment is #128:

50.r6: here Leela Zero #157 is the only AI that visited this joseki variation at the given
Opening more than one time:
59.f2: …and this time #157th became a singular part of Leela Zero family who rejected
this one space jump to the second line:

105.g8: all ELF OpenGo models can see this AGZ’s move, including the initial v0:
…at the same time, no MiniGo Network can see it, including the most recent v17:
This is the summary for matching the AlphaGo Zero 40B moves:

RANK AI’S RELEASE PTS MATCH FOUND


1 ELF OpenGo v1, 20B 2018-07 38 9/20 19/20
2 Leela Zero #204, 15B/40B 2019-02 37 9/20 18/20
3 ELF OpenGo v0, 20B 2018-05 36 7/20 19/20
4-7 Leela Zero #173, 20B 2018-09 34 8/20 18/20
PhoenixGo v1, 20B 2018-05 34 7/20 20/20
Leela Zero #188, 40B 2018-11 34 7/20 19/20
Leela Zero #226, 40B 2019-05 34 7/20 19/20
8-9 Leela Zero #174, 40B 2018-09 33 6/20 19/20
Leela Zero #224/ELF v2, 15B/40B 2019-05 33 6/20 18/20
10-11 MiniGo v15 #990, 20B 2019-01 31 6/20 17/20
ELF OpenGo v2, 20B 2019-02 31 5/20 19/20
12-13 MiniGo v17 #961, 20B 2019-03 30 7/20 16/20
Leela Zero #157, 15B 2018-07 30 5/20 19/20
14-15 Leela Zero #128, 15B 2018-04 28 5/20 17/20
MiniGo v16 #1144, 40B 2019-02 28 5/20 17/20
16 MiniGo v14 #979, 20B 2018-12 25 5/20 15/20
17 Leela Zero #116, 10B 2018-04 22 3/20 16/20
18 Leela 0.11 – checkpoint 2017-11 16 2/20 10/20

SOME OBSERVATIONS ON 40 BLOCKS PART

I) Observations on general compatibility with AlphaGo Zero’s moves

II) Observations on AI’s categories

III) Observations on AI’s families

RANK AI’S RELEASE PTS MATCH FOUND


1 ELF OpenGo v1, 20B 2018-07 75 18/40 38/40
2 Leela Zero #226, 40B 2019-05 72 17/40 36/40
3 PhoenixGo v1, 20B 2018-05 71 16/40 39/40
4 Leela Zero #224/ELF v2, 15B/40B 2019-05 69 15/40 36/40
5-7 Leela Zero #173, 20B 2018-09 68 16/40 35/40
Leela Zero #204, 15B/40B 2019-02 68 16/40 35/40
ELF OpenGo v0, 20B 2018-05 68 14/40 37/40
8-9 MiniGo v17 #961, 20B 2019-03 67 16/40 31/40
Leela Zero #174, 40B 2018-09 67 14/40 36/40
10 ELF OpenGo v2, 20B 2019-02 66 13/40 37/40
11 Leela Zero #157, 15B 2018-07 64 12/40 36/40
12 MiniGo v16 #1144, 40B 2019-02 60 13/40 33/40
13 MiniGo v15 #990, 20B 2019-01 59 13/40 31/40
14 Leela Zero #128, 15B 2018-04 57 11/40 34/40
15 MiniGo v14 #979, 20B 2018-12 50 11/40 28/40
16 Leela Zero #116, 10B 2018-04 49 8/40 32/40

+ГРАФИКИ МОДЕЛЕЙ ПО СЕМЕЙСТВАМ

S-ar putea să vă placă și