English version of Zen(ith) Go 6 (d’load, Windows, 7d)

[Mike Novack]

Lol. How come do specific time will make it stronger?

You are being told that there are two (different) ways to specify.

a) # of playouts ----- in which case the amount of time required for each move will depend on the hardware. This doesn't require the program to actually test what hardware it is running on, just takes as long as it takes. Increasing the number of playouts improves the play.

b) amount of time per move --- in which case, at the start, the program tests the hardware performance and decides how many playouts it can use to fit the time control selected. So on any given harware, increasing the time improves the play.

Note that this isn't linear. Below some number of playouts the program would be weak, erratic. And as the number of playouts increases, the amount of inporovement decreases (diminishing returns).

Let's note a few things:
a) The suggested 120 seconds per move is like OLD TIME speeds with games spread over a couple days.
b) The reason that the computer programs do better (than humans) at fast speeds isn't precisely the speed. AFAIK all the programs available lack a feature humans use --- decide that the position NOW is at a critical junction and use some reserved time to think. Or make a few real quick moves (wastes ko threats) to build up that reserve. It's an advanced "time management" problem. The point is, 10 second per move go is not quite the same as 360 moves in an hour even though the averae time per move is the same.

[kdbaby1412]

I totally see your point. It makes sense now.
When I put it at 7 dan, it takes less than 30 sec for a move.
But when I specify time at 30 sec, sometimes it used all 30 sec and most of the time it used more than 20 sec to play one move. And I do notice the difference in strength.

However, I won't say the program is as strong as 7d on KGS.

[jorden]

However, I won't say the program is as strong as 7d on KGS.

Do you give yourself the same time to play as the bot?

[pookpooi]

However, I won't say the program is as strong as 7d on KGS.

The 7d is from 20 cores cpu playing 15sec/move games. Most AI tend to win more in blitz.

[kdbaby1412]

However, I won't say the program is as strong as 7d on KGS.

Do you give yourself the same time to play as the bot?

No...I tend to play faster than the bot. Is this crucial to the bot strength?

[Mike Novack]

I totally see your point. It makes sense now.
When I put it at 7 dan, it takes less than 30 sec for a move.
But when I specify time at 30 sec, sometimes it used all 30 sec and most of the time it used more than 20 sec to play one move. And I do notice the difference in strength.

However, I won't say the program is as strong as 7d on KGS.

You have enough crunch with your 4790 that while maybe not 7d, probably at least 4-5. But I am curious what you mean, how you might be judging its strength if you are around 4 kyu.

How many stones do you have to take to beat it? Are you able to beat it, or are you thinking things like "well a stone is about 10 points, so if the program beats me by only 30 points it's only three stones better". That's the wrong way to think, especially with programs that make any use of MCTS.

[kdbaby1412]

I totally see your point. It makes sense now.
When I put it at 7 dan, it takes less than 30 sec for a move.
But when I specify time at 30 sec, sometimes it used all 30 sec and most of the time it used more than 20 sec to play one move. And I do notice the difference in strength.

However, I won't say the program is as strong as 7d on KGS.

You have enough crunch with your 4790 that while maybe not 7d, probably at least 4-5. But I am curious what you mean, how you might be judging its strength if you are around 4 kyu.

How many stones do you have to take to beat it? Are you able to beat it, or are you thinking things like "well a stone is about 10 points, so if the program beats me by only 30 points it's only three stones better". That's the wrong way to think, especially with programs that make any use of MCTS.

Ah I see. Thanks for the insight

How I test it? I let it played with my cousin, who is a 5d on KGS. He beat it about ~20 points. There are few mistakes that Zen make, but overall, I think it reaches high 4d on my computer.

[jorden]

No...I tend to play faster than the bot. Is this crucial to the bot strength?

Well, if you would play a human opponent, would it matter that they had more time or less time to play their moves?

[pookpooi]

Ah I see. Thanks for the insight

How I test it? I let it played with my cousin, who is a 5d on KGS. He beat it about ~20 points. There are few mistakes that Zen make, but overall, I think it reaches high 4d on my computer.

I'm not sure if one sample is enough to conclude, Hanayeol a go youtuber play against CrazyStone commercial version in 2 hours/side game and he lost all 5 games match. He's also 5D KGS. I think both CrazyStone and Zen are about the same rank, though right now Zen is much better in experiment version.

[Mike Novack]

How I test it? I let it played with my cousin, who is a 5d on KGS. He beat it about ~20 points. There are few mistakes that Zen make, but overall, I think it reaches high 4d on my computer.

But I will repeat, the margin of victory not a good indication of anything. All you know at this point is that with one game played, your 5d cousin won the game. You would know more after your cousin has played ten games against it.

[Gomoto]

I really like Zenith Go 6.

strong AI
clean user interface
good game review tool (Analyze function and moving through game with mouse wheel)
nice graphics (I would like to be able to turn off the small black border around the white stones)
fast startup
you can run two instances (to memorize pro games for example)
manual replay function (memorize pro games)
good sgf editor
fast startup
rated games
best of seven games
tsumego (300 problems)
voice over
cho chikun :-))))

Some things that could be improved:
add a tree view
option to move through game just with the cursor keys
improve movement through different variations

Zen is not only a good engine but also a very good go program.

[quantumf]

At a shade over $100 this is way too expensive, but just in case someone wants to buy me a Christmas present, any advice on how to register? It wants arcane zip codes, insists on selecting a state/prefecture from a drop down of Japanese strings, and wants telephone numbers in a specific format. Can I just put random numbers in?

Edit: the dlmarket link is far more reasonable.

[kdbaby1412]

So today, I let Zen play 3 blitz game with my cousin, 5D KGS. Zen win all 3 by resignation.

5s / move each.

zen strength really flourish during speed Go.
The program is run on normal priority under i7-4930k.

I did attach the file. Zen play W in all 3 games. Komi 6.5, japanese Rule. 1:00 minute start. Zen are set at 5s each. The opponent has 10 sec per Byo-yomi.

http://lifein19x19.com/forum/download/file.php?id=7937

http://lifein19x19.com/forum/download/file.php?id=7936

http://lifein19x19.com/forum/download/file.php?id=7935

[Satorian]

zen strength really flourish during speed Go.
The program is run on normal priority under i7-4930k.

Zen generally runs better when set to a specific thinking time rather than a specific strength especially on fast CPUs, as it then utilizes all cores as long as possible instead of a single core until a preset calculation limit is hit.

[kdbaby1412]

Indeed true. I found this is very interesting. I will test out more game.

So far, on my CPU, Zen play 6 games with 5d KGS and manage to win 5