Who wants a pro player at home for free ?

[Vargo]

In answer to a question on GitHub, AQ's author said :

...
AQ has been officially running at Yugen no ma, which is a Go server managed Nihon Ki-in, since last week and the result of 100 games against Japanese professional players was 95-5 (win-loss). The hardware is dual GPUs (GTX 1080Ti) and i7-8700K. If you use a single GTX 1080Ti, the elo rating will be reduced by 100-120.

I knew that with a good GPU, AQ was very strong, but 95% wins against japanese pros with 2 GPUs is really amazing. And only 100-120 Elo less with 1 GTX 1080Ti... that's your pro at home

[Uberdude]

When it played in that world computer go tournament in Mongolia I thought AQ played quite a nice style, and initially held its own against FineArt, Zen, CGI before messing up when things got too complicated. In these games versus (mostly junior) Japanese pros it still seems a fairly clean and unremarkable style, but better than before. I saw one win vs Shibano Toramaru 7p (probably strongest youngster, now in goratings top 100) by 7.5 points.

[Vargo]

I bet FineArt or DeepZen run on much more powerful computers. What I find truly amazing is achieving pro strength with a standard home computer with a good GPU.

For me, there's only one feature missing... AQ doesn't handle more than 3-4 stones handicap, so I can't play against it, as I would probably need 9 stones

But the author said that in the next update, AQ will be able to play high handicap.

Alphago, AQ, Leela Zero, what a blessed period for computer go !

[Gomoto]

AQ GPU versus Zenith 7 CPU, ongoing ...

[Gomoto]

At 20 sec per move it is not a pro player at home yet

AQ GPU black (Nvidia 1080 Ti) vs Zen 7 white (CPU Intel I5 2500k)

0 : 1

Zenith 7 won by resign, not 2 bad for CPU bound Zen 7

(Move 129 by AQ at N11 was probably the loosing move, should have been probably at P16 to keep a chance)

[EdLee]

AQ @ Nihon Ki-in:

[Vargo]

At 20 sec per move it is not a pro player at home yet
AQ GPU black (Nvidia 1080 Ti) vs Zen 7 white (CPU Intel I5 2500k)

There's obviously something wrong here...
It's not possible to have a demonstrable 95% win ratio against pros and lose against Zen.

Do you have the full AQ version ?
Does AQ correctly detect and use your graphic card ? (it should be red hot !)
Do both programs have comparable thinking time ?
Is your config file Ok (enough threads for example)
is pondering off for both ?

I don't have Zen 7, but I have Zen 6, and some months ago, I tried some AQ2.0.3 - Z6 games. I don't remember the exact results, but it was similar to the games with CSDL, there was at least 2-3 stones difference at something like CGOS settings. Games against CSDL or ZEN can't be played in Sabaki, you have to enter every move by hand, not a great incentive to run a lot of these matches...

Anyway, I don't think I can convince you, but just look at AQ's DEMONSTRABLE results, obtained with a simple gaming PC, so yes, with 100 Elo less than a 95% winner against pros, you're probably still in the pro zone .

[Gomoto]

No worries, I think AQ is a very strong bot.

I just wanted to run a test for fun to see if Zen 7 is still a match on my computer. And I was nicley surprised that CPU based Zen 7 on a 2500K@4.2 Ghz can still beat AQ GPU on a 1080 TI (at least once ). My only conclusion: The policy net of Zen 7 seems to be of high quality.

(By the way, one should not underestimate Zen 7 by looking at some pro playing on youtube versus Zen 7 on any mobile cpu. Zen 7 on a real desktop CPU is a different beast.)

AQ GPU 2.1.1
pondering is off for both bots
both bots have 20 sec per move (AQ spending 20 sec every move, Zen7 is not using the 20 secs for every move)

[pookpooi]

At 20 sec per move it is not a pro player at home yet
AQ GPU black (Nvidia 1080 Ti) vs Zen 7 white (CPU Intel I5 2500k)

There's obviously something wrong here...
It's not possible to have a demonstrable 95% win ratio against pros and lose against Zen.

First, beside from different number of GPU , CPU are also different (i5 vs. i7)
Second, need many many more games. One game is never enough in any statistics.

[Mike Novack]

First, beside from different number of GPU , CPU are also different (i5 vs. i7)
Second, need many many more games. One game is never enough in any statistics.

And not JUST the i7 vs i5 but a specific VERY high crunch power i7 was specified (there is a significant range in crunch power within the i7 family). We need to keep in mind that the performance of these programs can be very non-linear (given the need to move withing a defined amount of time).

The claim should be properly tested, can this program perform at the claimed level given an i7-8700K (or something of equivalent crunch power) plus a GTX 1080Ti (or equivalent). Important because at the level of a high end "gaming machine" or high end workstation.

[Calvin Clark]

Free? This must refer to the period between buying such hardware and getting your next power bill.

[Gomoto]

AQ GPU takes revenge at 30 seconds per move

AQ versus Zen
1 : 1

[Mike Novack]

Free? This must refer to the period between buying such hardware and getting your next power bill.

Maybe look up? While a i7-8700K and 1080Ti too much for a laptop, not untypical for a desktop.

BTW, something like this, significantly better performance when given 50% more time is consistent with it just not having enough crunch power to play at 20 sec/move. The 8700K would have about 50% more crunch power than the previous high end i7's

[hydrogenpi7]

Free? This must refer to the period between buying such hardware and getting your next power bill.

Maybe look up? While a i7-8700K and 1080Ti too much for a laptop, not untypical for a desktop.

BTW, something like this, significantly better performance when given 50% more time is consistent with it just not having enough crunch power to play at 20 sec/move. The 8700K would have about 50% more crunch power than the previous high end i7's

AQ on GTX 970 just beat a 4P tonight.
Just wait until GTX 2080 comes out next month....
Humans dont stand a chance anymore.

Pretty soon everyone can have access to "ALphaGO"