Search found 244 matches

by MichelG
Fri Aug 14, 2015 15:29
Forum: Draughts, Computer, Internet
Topic: Tournament of programs in international draughts
Replies: 25
Views: 11359

Re: Tournament of programs in international draughts

I am Awaiting the results...

Will you play all games with the damexchange prototcol?

Michel
by MichelG
Thu Aug 13, 2015 12:56
Forum: Draughts, Computer, Internet
Topic: Eval tuning
Replies: 41
Views: 21188

Re: Eval tuning

I meant the output that numerical optimizers usual give, so for each iteration in their solver they print out the objective function (sum of squares, log-likelihood, whatever), the gradient and/or the step size. If you would use the Scan eval() and feed it your training positions, and evaluate your...
by MichelG
Thu Aug 13, 2015 09:14
Forum: Draughts, Computer, Internet
Topic: Eval tuning
Replies: 41
Views: 21188

Re: Eval tuning

Very interesting experiments, Michel! I am just curious how much of a better fit the Dragon-ML eval is compared to Scan. So if you would run the Scan eval() on your 900M training set, how well does it predict the game scores compared to your own 50M-parameter eval()? Not sure if you do log-likeliho...
by MichelG
Wed Aug 12, 2015 20:18
Forum: Draughts, Computer, Internet
Topic: Eval tuning
Replies: 41
Views: 21188

Re: Eval tuning

I have integrated the scan evaluation function into dragon for testing purposes and done some testing. I have tested against 2 different eval function of dragon: 1) DragonBS This evalution function is the original eval function of dragon. It is a manual tuned "classic" eval. It includes a breakthrou...
by MichelG
Mon Aug 10, 2015 08:35
Forum: Draughts, Computer, Internet
Topic: dragon 4.5
Replies: 52
Views: 30994

dragon 4.5

Amidst other stuff i am working on, there is a new version available for dragon draughts. Version 4.5 can be downloaded from: http://mdgsoft.home.xs4all.nl/draughts/download/SetupDragon%20v4.5.0.msi Improvements include - minor strength gain - better support for 4k monitors - various bugfixes, inclu...
by MichelG
Sat Aug 08, 2015 00:23
Forum: Draughts, Computer, Internet
Topic: Eval tuning
Replies: 41
Views: 21188

Re: Eval tuning

Dragon uses conjugated gradient algorithm for doing that. I found that in the 'numerical recipes in c' book. I made it run multi-threaded and minimized the memory footprint. Gamefase 3 (the endgame) is currently the biggest and learned from 900 million examples to learn 17 million parameters. This t...
by MichelG
Fri Aug 07, 2015 20:23
Forum: Draughts, Computer, Internet
Topic: Internet engine matches
Replies: 470
Views: 182200

Re: Internet engine matches

Thanks Ed and Michel! While I'm not complaining by any means, those results lead to a puzzle for me. Dragon Draughts already has the most accurate evaluation (through years of effort) and it's also gaining a lot from a breakthrough table that Scan doesn't have. I don't believe that my (very naive) ...
by MichelG
Fri Aug 07, 2015 12:42
Forum: Draughts, Computer, Internet
Topic: Internet engine matches
Replies: 470
Views: 182200

Re: Internet engine matches

I have run a engine match between dragon v4.5 and scan with similar parameters as ed: - 4 threads - scan: 6 piece databases, 256 mb hash - dragon: selected 8 piece databases, 6x6 eval patterns - 3 minutes for 75 moves I used 3 move ballots. The match is still running, but the first 158 games ended i...
by MichelG
Thu Aug 06, 2015 22:24
Forum: Draughts, Computer, Internet
Topic: Internet engine matches
Replies: 470
Views: 182200

Re: Internet engine matches

I ran an engine match of 2-move start positions, kingsrow vs. scan, each using 6-piece dbs, 4 search threads, 120 moves in 3 minutes time controls. Results from kingsrow's perspective: +0, -33, =125, unknowns 0 Scan is really strong! Congrats to Fabien for creating a very powerful draughts program,...
by MichelG
Thu Jul 30, 2015 09:32
Forum: Draughts, Computer, Internet
Topic: Internet engine matches
Replies: 470
Views: 182200

Re: Internet engine matches

I wonder what the limits of machine learning are. It seems ideally suited to exploit local patterns (in particular locks, perhaps breakthroughs), but global concepts (total terrain advancement, tempo development etc.) seem hard to fit with overlapping patterns. What are your thoughts on it? If you ...
by MichelG
Thu Jul 30, 2015 09:24
Forum: Draughts, Computer, Internet
Topic: Perfect Play
Replies: 9
Views: 6321

Re: Perfect Play

see my post in the Matches thread. I don't think programs are near perfect play because the drawing margin (438 ELO points!) is too high and shields a lot of subtle positional errors. All the high drawing rates proves is that most programs have similar knowledge and search and are more or less equa...
by MichelG
Sat Jul 25, 2015 12:26
Forum: Draughts, Computer, Internet
Topic: Protocols and tools
Replies: 42
Views: 23165

Re: Protocols and tools

I think Bert is right in that the draughts community is very small, and (a long time ago) i was happily surprised to me that even something like DXP even exists. At the lowest level dragon uses memory mapped files and events to communicate between the gui and the engine. It still should support sock...
by MichelG
Sun Jul 12, 2015 19:53
Forum: Draughts, Computer, Internet
Topic: Dragon 4.4.0
Replies: 2
Views: 2870

Re: Dragon 4.4.0

i haven't seen this error before, but i will look into it.
Michel
by MichelG
Sat Jul 11, 2015 21:22
Forum: Draughts, Computer, Internet
Topic: reproducing scan's eval function
Replies: 3
Views: 3388

Re: reproducing scan's eval function

Dragon uses the conjugated gradient algorithm for computing the weights. It's pretty fast for 1 million positions, taking maybe 30 seconds or so. But you can imaging playing a million games takes a lot longer than that. I matched dragon with the newly learned weights and the scan eval, both with dra...
by MichelG
Fri Jul 10, 2015 21:21
Forum: Draughts, Computer, Internet
Topic: reproducing scan's eval function
Replies: 3
Views: 3388

reproducing scan's eval function

Today, i used dragon to recreate the eval function of scan, using dragon's implementation. For the test, i generated about 1 million games, and used the learning technique to learn the same 16 8-field patterns as scan does. Some differences though: - dragon patterns includes kings. - dragon uses 4 g...