There is a new release MS-169d ready for download at http://www.10x8.net/down/SmirfEngine.dll for owners of a SMRF key set. It still has no opening library, but it should again be marginally improved.
It would be fine to know, which programs at which release stage are at the start now. And it would be nice to read some details on its authors ...
Subscribe to:
Post Comments (Atom)
17 comments:
The Martillo Program is ready to play. I can describe it briefly
64 bit Gothic Chess program
Minimal window PVS search
Automatically tuned evaluation function created from 10,000 games of a.i. versus a.i games run on 100 servers over 10 days (10 games per day)
Plays more for positional gains than tactical gains
Sparring partners were Gothic Vortex copper and recompiled ChessV to remove Gothic Chess restriction.
Search speed = 1 million positions per second on fastest AMD machines out now. Probably ~700,000 per second on machine for tournament.
And program authors may email me at this address (to avoid mixing school and work with programming talks)
MartilloProgram@gmail.com
Good to see you here Manuel, and thanks for entering your program in the tournament. I'd like to talk to you more about the tuning procedure you used. It sounds like an interesting project!
Also, to other programmers, just a reminder, I MUST HAVE your programs no later than Monday afternoon for this week's battles! Any program submitted after 12 noon EASTERN time on Monday will not be eligible to play in round 1.
Programmers may submit new programs once per wek, if need be, for the duration of the tournament, but only those received on the nearest Monday will play in the upcoming cycle of games.
Welcome to the Gothic Chess world martillo, and good luck in the tournament! Your program sounds impressive and I think you will be keeping the other guys busy this time around.
So is Martillo a free program? Where can I get it if so. Thanks!
GothicChessInventor:
Good to see you here Manuel, and thanks for entering your program in the tournament. I'd like to talk to you more about the tuning procedure you used. It sounds like an interesting project!
Yes it sounds an interesting project but....
A big BUT: If you know about all these, and you know, then you should know by now that hand tuned values are always better than any kind of values produced by automatic evaluation procedures, with any method!
Doesn't seem logical but it is true!
All these methods for finding "best" values are damn interesting, but i would not spend even a second since just good guesses would perform better :-)
About program tuning:
More than once I adjusted the piece weights to Gothic Vortex. I sent test versions to Rob Colanzi, better known as ChessCarpenter on BrainKing and GothicChessLive.
Whenever I thought I had some "really good" values that should hypothetically be better, Rob would tell me he played engine-vs-engine matches, and my old weights were invariably better.
It took me over 2 years to produce a version that would defeat the old Vortex, and this required introduction of new positional knowledge, better search extension code, a compeltely new king safety model, and not better piece values.
So, I have always wondered:
Could some mechanical engine-vs.-engine match where the piece values were adjusted produce an "ultimate" function when used with this new positional knowledge?
I think it is fair enough to say that manually tuning a program is time consuming and difficult. If you stratify the position code from the material evaluation, and only auto-tune the material evaluation, you stand a much better chance of improving the program overall.
If one is looking at a function giving a quality index for a given combination of evaluation parameters, you would see an uncontinuous multidimensional object for optimizing. It would be full of myriads of local optima. To find something real better thus does not mean simply to find a very good combination of parameters, instead it would be a much bigger step to change the underlaying set of evaluation components to support a more continuous optimization object.
A quote from a Deep Thought team member is
We observed that deeper searches in the tuning code lead to better results, even though it could be argued that evaluations should be orthogonal to searching. Perhaps an explanation for this effect is that deeper searches lead to more differences in the positions that are being related because they are more moves apart. Therefore, the tuning process collects more information on how individual components of the evaluation relate to each other. For example, the tuning process did result in a better understanding of the piece-values with respect to each others and as a function of the amount of material left on the board (this was used to control the transitions from opening, mid-game to end-game).
A. Nowatzyk
I read their paper and did something similar.
I remember reading something about the Deep Thought team abandoning their automated tuning evaluation function when it was underperforming in actual play. It's been so long, perhaps I don't have a clear picture of the whole story.
So is my favorite program gonna win, smirfy?
:)
:)
:)
Hi LeGuSwi, remember, that SMIRF will be playing here as a zombie, because it had been terminated some months ago and now has been revitalized by some 'magic' improvements for this tournament. But though I am just working on one more new miraculous improvement, I still would be very happy, if SMIRF could manage to become second best in the end. But maybe, its fans could chearlead it through all obstacles ... ;-)
SMIRF's next release MS-169e is now ready for download again at http://www.10x8.net/down/SmirfEngine.dll for owners of a SMIRF key set only like the GCInventor. It still has no opening library, but again it should be a little bit improved. It is intended to be used having selected a cache size of 256 MB. SMIRF's known weaknesses are to sufficiently use pawns to perform a good opening and to mate the opponent (if it has the chance to) via the shortest combination.
We're getting closer to the deadline for the final submission for round 1 of the tournament.
Monday at NOON eastern I need them all so I can install them on the machine and bring it to the playing facility which is over an hour from my house.
Good Luck to all the Gothic Chess programs this year.
This should be and exciting year with many strong programs!
I will be on hand tonight at the MainLine CC to see the 1st round which is Martillo vs. Gothic Vortex!
Go Gothic Chess
Post a Comment