Computer Olympiad 2015 Leiden

[BertTuyt]

It seems that during the Computer Olympiad the Draughts tournament will have most participants.

Here is the list so far:
Sjende Blyn, Moby Dam (formerly knows as Dam 2.X), Tornado, Deep Nikita, Damage, Dragon Draughts, Cerberus, Scan (author Fabien Letouzey, famous in computer chess!), GWD, JDraughts, Slagzet.com and Tdking.

Missing in action unfortunately (so far) Damy, Maximus, and Kingsrow.
If we would have these 3 programs also, it would be without doubt the strongest Draughts competition ever!
So I hope on good news in the next weeks.

I can understand that it is not easy for Ed to fly to the Netherlands, but an option could be that Rein operates KingsRow, from my perspective that would be really great.

Last but not least, I still prefer a double round tournament, and if we limit the thinking time to 15 mins each side, and we play 2 consequetive games (with changed colors) than we could manage that in 2 days.

Bert

[Rein Halbersma]

It seems that during the Computer Olympiad the Draughts tournament will have most participants.

Here is the list so far:
Sjende Blyn, Moby Dam (formerly knows as Dam 2.X), Tornado, Deep Nikita, Damage, Dragon Draughts, Cerberus, Scan (author Fabien Letouzey, famous in computer chess!), GWD, JDraughts, Slagzet.com and Tdking.

Missing in action unfortunately (so far) Damy, Maximus, and Kingsrow.
If we would have these 3 programs also, it would be without doubt the strongest Draughts competition ever!
So I hope on good news in the next weeks.

I can understand that it is not easy for Ed to fly to the Netherlands, but an option could be that Rein operates KingsRow, from my perspective that would be really great.

Last but not least, I still prefer a double round tournament, and if we limit the thinking time to 15 mins each side, and we play 2 consequetive games (with changed colors) than we could manage that in 2 days.

Bert

I am on holiday in that period and unable to attend (already booked before I knew about the tournament). Of course there are many experienced Kingsrow users that Ed could try to enlist if he would not come himself.

[Eric van Dusseldorp]

When is the tournament?

[Ed Gilbert]

Here is an email I received in March that has some details of the tournament:

L.S.

For those who forgot that there is a computerdraughts tournament during the Mind Sport Olympiad in July 2015:

Days of the tournament : 4 and 5 July 2015
Location : Leiden University in Leiden The Netherlands (near The Hague).

Up to now we have the following participants:

Cerberus, Leo Nagels
Dam 2.2, Harm Jetten (probably a different version number)
GWD, Klaas Bor, Gijsbert Wiesenekker
TDKing, Ton Tillemans
Tornado, Frank Mesander

It would be nice if we would have some additional programs in the tournament for more excitement.

If you want to join the list of participants, please answer this e-mail.
If you won’t join for one reason or another, please answer this e-mail as well.

You find the link to the official subscription form at
https://docs.google.com/forms/d/1eMMw5- ... rm?c=0&w=1
Don’t forget that a contribution fee of 50 Euro is due.

General information on the tournament is at http://www.icga.org

Friendly greetings,


Leo Nagels

I am unable to attend. If anyone is interested in attending as an operator for kingsrow's entry, please email me. I would of course take care of the entry fee.

-- Ed

[Harm Jetten]

Hello all,

The unofficial results of the Computer Olympiad 10x10 draughts tournament are available on my web page.
Also I am providing information on my new program/engine Moby Dam, including source code.
Enjoy!
Harm.

http://hjetten.home.xs4all.nl/dameng.html#perform

[Rein Halbersma]

Hello all,

The unofficial results of the Computer Olympiad 10x10 draughts tournament are available on my web page.
Also I am providing information on my new program/engine Moby Dam, including source code.
Enjoy!
Harm.

http://hjetten.home.xs4all.nl/dameng.html#perform

Hello Harm,

Congratulations on your new program's result. And very nice to publish its source! I will study it

[H.G.Muller]

So how does the Draughts-programming community look upon the victory of Scan? Is this a sensational and revolutionary development? I understood Fabien's program did not even use end-game tables.

[MichelG]

So how does the Draughts-programming community look upon the victory of Scan? Is this a sensational and revolutionary development? I understood Fabien's program did not even use end-game tables.

Congratulations to Fabien!

Scan's result has been excellent. The reason it is so strong seems that it is using machine learning techniques. I don't know any other program apart from dragon draughts that uses large scale machine learning. Scan and dragon use similar learning techniques

I think for a top draughts program today, you need multithreading, learning, endgame databases and the latest game search techniques. Scan is using them all, and has apparently implemented them very well. (scan did use 6 piece endgame tables)

It is a good lesson for all draughts programmers out there, and it is taking us 1 step closer to near-perfect play in 10x10 draughts.

Michel

[Krzysztof Grzelak]

Congratulations to Fabien!

Please write as someone knows what versions of the programs participated in the Olympics.

[Rein Halbersma]

So how does the Draughts-programming community look upon the victory of Scan? Is this a sensational and revolutionary development? I understood Fabien's program did not even use end-game tables.

Congratulations to Fabien!

Scan's result has been excellent. The reason it is so strong seems that it is using machine learning techniques. I don't know any other program apart from dragon draughts that uses large scale machine learning. Scan and dragon use similar learning techniques

I think for a top draughts program today, you need multithreading, learning, endgame databases and the latest game search techniques. Scan is using them all, and has apparently implemented them very well. (scan did use 6 piece endgame tables)

It is a good lesson for all draughts programmers out there, and it is taking us 1 step closer to near-perfect play in 10x10 draughts.

Michel

What kind of machine learning? Offline fitting of eval coefficients and/or book expansion? Or online learning of dynamic search parameters?

[MichelG]

What kind of machine learning? Offline fitting of eval coefficients and/or book expansion? Or online learning of dynamic search parameters?

Offline learning of Eval coefficients. Whereas a human can fine-tune tens or may a hundred eval parameters, machine learning takes it a step further. Scan was using 50000 or so eval parameters, dragon used 49000000.

Michel

[Fabien Letouzey]

It is a good lesson for all draughts programmers out there, and it is taking us 1 step closer to near-perfect play in 10x10 draughts.

Hi all,

I will try to address topics separately. Maybe we can start by having a look at Scan's games. I add that Scan did not play any serious game prior to the tournament, so those 11 are the only ones we have. Someone might want to sanitise the PDN file. It's in Unix text format, the tags were edited manually and are incomplete, and the program names come from the tournament pairing sheet. For example I think that Slagzet should actually be Slagzet.com, etc...

While operating I could see only one terrible blunder, in the game with JDraughts. Someone can post a diagram or something: Scan let JDraughts place a king-making combination early in the game. Scan did not consider the king to be so dangerous because it did not have any safe move. What Scan did not know however is that it eventually (as in "beyond the horizon") had to let it roam free. The evaluation learning is not at fault though, because kings are scored "manually" and I more or less neglected that part as a first approximation. JDraughts did not manage to win this game, but it seems logical to think that there must have been (at least) an opportunity somewhere ... Scan's evaluation for most of the game was -0.50 (starting a few plies after the blunder), JDraughts was also stable, but at +3.00.

Fabien.

[Rein Halbersma]

It is a good lesson for all draughts programmers out there, and it is taking us 1 step closer to near-perfect play in 10x10 draughts.

Hi all,

I will try to address topics separately. Maybe we can start by having a look at Scan's games. I add that Scan did not play any serious game prior to the tournament, so those 11 are the only ones we have. Someone might want to sanitise the PDN file. It's in Unix text format, the tags were edited manually and are incomplete, and the program names come from the tournament pairing sheet. For example I think that Slagzet should actually be Slagzet.com, etc...

While operating I could see only one terrible blunder, in the game with JDraughts. Someone can post a diagram or something: Scan let JDraughts place a king-making combination early in the game. Scan did not consider the king to be so dangerous because it did not have any safe move. What Scan did not know however is that it eventually (as in "beyond the horizon") had to let it roam free. The evaluation learning is not at fault though, because kings are scored "manually" and I more or less neglected that part as a first approximation. JDraughts did not manage to win this game, but it seems logical to think that there must have been (at least) an opportunity somewhere ... Scan's evaluation for most of the game was -0.50 (starting a few plies after the blunder), JDraughts was also stable, but at +3.00.

Fabien.

Congratulations on your win, Fabien!

The entire tournament should be on toernooibase soon, which makes it easier for people to analyze them. I'm currently on vacation and without decent computer access but I'll try and analyze the games this weekend.

BTW, is your engine Scan going to be open source? I'd love to learn your approach to the eval in full detail.

[Fabien Letouzey]

BTW, is your engine Scan going to be open source? I'd love to learn your approach to the eval in full detail.

Scan will be open source but this will only include code directly related to game playing. The development version is more like a library of modules to allow me to run "lab" experiments.

Furthermore the learning stuff is very old. Regression techniques for position evaluation date back to 1982 apparently (http://chessprogramming.wikispaces.com/ ... Regression).
In the case of 10x10 draughts, Dragon already used evaluation learning about 8 years ago IIUC. I have not found evidence of a public discussion though.

I use Othello stuff in Scan, so you should re-read the papers by Michael Buro.
They are all good but you seem in a hurry so I suggest starting with "Statistical Feature Combination for the Evaluation of Game Positions" for an accessible overview and "Experiments with Multi-ProbCut and a New High-Quality Evaluation Function for Othello" for mathematical details.

[Krzysztof Grzelak]

MichelG which version you played at the Olympics.