ymgaq released Windows version of his AQ v2.0.3 bot

aTan · #1

https://github.com/ymgaq/AQ/releases

- Make it compatible with Windows (AQ_win)
- Added AQ-mini (AQ-mini_win: for CPU only)
- Supports 3 or more multi-GPUs
- Added Ponder OFF setting

pnprog · #2

I just want to report that I ran AQ vs Leela games on my computer, alternating black and white for both bots.
Both bots were running CPU only (no pondering), and my CPU is pretty old one: i3-3217U CPU @ 1.80GHz (4 cores).
Time settings was 30 second by moves, which AQ used at the fullest, but Leela somehow only used half of its time.

Result:

AQ: 10 wins
Leela: 4 wins

On shorter time (10 s/move) Leela seems to win all the games.

AQ is definitively a strong bot. I have added support for it in Go Review Partner, but there are still some details to be figured out with the developer of AQ (namely how to understand AQ win rate values).

luigi · #3

It doesn't work for me. Is it 32-bit or 64-bit? Do I need to enter any commands on Sabaki?

pnprog · #4

Hi!

luigi wrote:

It doesn't work for me. Is it 32-bit or 64-bit? Do I need to enter any commands on Sabaki?

The easiest way for you to check that your computer can run AQ is to run directly the executable file AQ.exe, and once the black terminal looking windows is opened, wait until you see the message configuration loaded. to input genmove black followed by a press on <enter> key.

AQ will start thinking then provide an answer, it will look something like this:

Code:

configuration loaded.
genmove black
thinking...
move cnt=1: left time=0.0[sec]
5[nodes] 5.1[sec] 685[playouts] 134.5[pps/thread]
total games=685, evaluated policy=5(0), value=23(24)
|move|count  |value|roll |prob |depth| best sequence
|D4  |    189| 51.9| 45.5| 17.5|    1| D4 ->Q16
|Q4  |    141| 51.5| 44.0| 18.0|    1| Q4 ->O3 
|Q16 |    109| 51.3| 36.7| 18.7|    1| Q16->D16
|D16 |     73| 50.7| 37.0| 17.3|    1| D16->F17
|Q17 |     53| 51.5| 54.7|  3.9|    1| 
|C16 |     25| 49.8| 52.0|  3.4|    1| 
|Q3  |     22| 51.4| 45.5|  3.3|    1| 
|C4  |     19| 50.6| 42.1|  3.8|    1| 
|R16 |     19| 51.5| 36.8|  4.0|    1| 
|D17 |     17| 50.5| 41.2|  3.4|    1| 
= D4

   A  B  C  D  E  F  G  H  J  K  L  M  N  O  P  Q  R  S  T 
19 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 19
18 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 18
17 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 17
16 .  .  .  +  .  .  .  .  .  +  .  .  .  .  .  +  .  .  . 16
15 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 15
14 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 14
13 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 13
12 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 12
11 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  . 11
10 .  .  .  +  .  .  .  .  .  +  .  .  .  .  .  +  .  .  . 10
 9 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  9
 8 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  8
 7 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  7
 6 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  6
 5 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  5
 4 .  .  . [X] .  .  .  .  .  +  .  .  .  .  .  +  .  .  .  4
 3 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  3
 2 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  2
 1 .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  1
   A  B  C  D  E  F  G  H  J  K  L  M  N  O  P  Q  R  S  T 

(here the answer is D4)

At that point you can be confident that AQ is working on your computer.

luigi · #5

pnprog wrote:

luigi wrote:

It doesn't work for me. Is it 32-bit or 64-bit? Do I need to enter any commands on Sabaki?

The easiest way for you to check that your computer can run AQ is to run directly the executable file AQ.exe

I just tried that and got the following message (translated from Spanish by me):

Quote:

This version of AQ.exe is not compatible with the Windows version that you are running. Check out the computer system information to see if you need an x86 (32-bit) or x64 (64-bit) version of the program, and then contact the software editor.

Is your computer 64-bit?

pnprog · #6

luigi wrote:

Is your computer 64-bit?

Yes, 64 bits operating system, x64 processor.

Vargo · #7

It works fine with Win10
install Sabaki
manage engines...
Add AQ
Add Leela replace "No argument" by "--gtp --noponder"
in AQ configuration file, -self match =off and -use pondering =off
and that's it, in sabaki, File, New, you choose your black engine, your white engine, and the game begins you can see at the same time the graphic board and the information (winrate etc)
Really fascinating...

THANK YOU to the authors of AQ, Leela and Sabaki !!!

luigi · #8

Vargo wrote:

It works fine with Win10

I'm on Windows 7. I wonder if that's the problem. But it's more likely the 64-bit thing. My computer is 32-bit.

Quote:

install Sabaki
manage engines...
Add AQ
Add Leela replace "No argument" by "--gtp --noponder"
in AQ configuration file, -self match =off and -use pondering =off
and that's it, in sabaki, File, New, you choose your black engine, your white engine, and the game begins you can see at the same time the graphic board and the information (winrate etc)

Leela works, but AQ doesn't. :sad:

tartaric · #9

Can someone explain how to make this bot working on Linux please? Can you explain every steps cause I am a noob using it. Thanks.

splee99 · **#10**

Well, without GPU on my computer, AQ really can't handle the latest Leela 0.11.0. I run twogtp and gogui and time setting is about 15 seconds per move. Leela 0.11.0 is too strong.

Vargo · **#11**

I think you are right, AQ needs a good GPU to be strong.

In Crazy Stone Deep Learning, you set the level at 13k, 12k, ... up to 6 Dan and 7 Dan , a very nice feature.
It's a way to play consistently at a certain level, independently of the hardware.

I set up 4 games between AQ (20" per move) and CSDL 5 Dan (6 dan and 7 dan take some time, and I'm a bit impatient !)
Both programs were side by side, and I entered the moves "by mouse" alternatively, no GTP

H3 : CSDL5dan - AQ(w) AQ wins
H4 : CSDL5dan - AQ(w) AQ wins
H5 : CSDL5dan - AQ(w) AQ wins
H6 : CSDL5dan - AQ(w) CSDL wins

In the H5 game , AQ had a rather strange beginning, but managed to kill nicely a big black group.

Two questions :
1) Could someone (with a decent GPU and CSDL and AQ installed) set up something similar ?
2) What do you think it says about AQ's strength ?

I've used Sabaki for AQ's moves, without any argument.

AQ's configuration :

### Hardware setting ###
# Set number of 1-8 for 'gpu count'.
# 'thread count' is number of the total usable threads.
-gpu count =1
-thread count =8

### Time contorol setting ###
# If using 'time_left' command via GTP,
# set 'time controll' to 'off'.
-main time[sec] =0
-byoyomi[sec] =20
-time controll =on

### Thinking setting ###
-japanese rule =off
-komi =7.5
-symmetrical index =0
-mimic go =off
-never resign =off

### Others ###
-self match =off
-save log =on
-master =off
-worker =off
-pb path =pb
-resume sgf path =
-worker count =1
-use pondering =off

splee99 · **#12**

I just upgraded my computer with a EVGA GTX 1050 video card. After run some games between AQ and Leela, I found AQ is indeed very strong. I set the time to 10 sec per move for AQ and no time limit for Leela (usually it uses 15 sec per move). AQ is able to beat Leela 0.11 with 2 handicaps. Also it appears that you have to use --nobook option for Leela to let it think more thoroughly in the beginning.
Maybe the strength of AQ comes from its large memory usage, which is about 3.8 Gbytes when it is thinking, compared to 80 Mbytes for Leela. Also the number of threads used by Leela can't be set higher than 2.

kdbaby1412 · **#13**

Run this AQ on my computer with Leela and test how these two are fight together.

First game: AQ (Black) vs Leela 11.0 (White) - B+Res (1:0)
Second game: AQ (White) vs Leela 11.0 (Black) - B+Res (1:1)
Third game: AQ (Black) vs Leela 11.0 (White) - B+Res (2:1)

Computer spec: AMD Ryzen 1700X, GTX 1060 3GB

Vargo · **#14**

It seems very odd. .. I don't believe leela can win an even game.
You have 8 cores, how many threads do you have in the AQ config file ?
Pondering is off for both ?

kdbaby1412 · **#15**

Vargo wrote:

It seems very odd. .. I don't believe leela can win an even game.
You have 8 cores, how many threads do you have in the AQ config file ?
Pondering is off for both ?

First game I switch pondering on, and Leela was unable to do anything.

Second and Third I turned pondering off.

All game only used 4 threads. I'm currently testing with 8 thread right now. AQ is kinda slow even though setting is 5s/move byo-yomi.

I thought AQ used sole GPU and does not rely on CPU that much?

Mike Novack · **#16**

kdbaby1412 wrote:

AQ is kinda slow even though setting is 5s/move byo-yomi.

What do you mean by that? These programs probably have "time management" and ideally begin by analyzing the speed of the hardware on which they are running and setting themselves to utilize as much time as they have according to the time control setting.

Of course strength settings may play a role too, so if being asked to play below maximum possible strength might do that in a way that resulted in moving faster (using less time than the time control allowed).

In other words, by my way of thinking, if you tell a program "you can use 5 seconds per move" I would not call it "slow" if it used 4.5 sec/move. You said it could.

kdbaby1412 · **#17**

That's make sense.
What I mean is in its config file, I listthe byo-yomi time as 5 sec, and main time is 0. So I assume it can used 5 sec

kdbaby1412 · **#18**

Here game 4 and 5

Game 5, AQ (W) give Leela 11.0 2 stones handicap.

For some reason, in game 5, both bots do not pass and just keep playing in its territory. Chinese rule: Leela +1.5; Japanese rule: AQ wins

Vargo · **#19**

Quote:

All game only used 4 threads.

With an i7 quadcore, AQ's author recommends 5 threads, you should probably use more than 4.

Or maybe AQ is much weaker at 5" ?? For my 10 games, I had AQ and Leela both at 30" per move (time_settings 0 30 1)
no pondering

Results were :
Even games : AQ beats Leela0110GTP_OpenCL 5-0
H2 : AQ (W) beats Leela0110GTP_OpenCL (B) 4-1 (always by resignation)

My computer : Windows 10 i7-6700K (4gHz) 32GB GTX1080
(Leela benchmark : 200000 games in 9.27s. 21574 g/s 2696 g/s per thread)

If you have CSDL installed, I'd be very much interested in some H5 games against CSDL set at 5dan , because I believe that CSDL-5dan doesn't depend much on the hardware, it could even be a kind of reference point.

splee99 · **#20**

Does AQ know double ko?

Maybe not. This gave leela 0.11.0 a lot of advantage. Time setting is 15 seconds per move for each player. The GPU is EVGA GTX 1050.

ymgaq released Windows version of his AQ v2.0.3 bot

Who is online