Thank to Harm for his interesting automated tournament site!
Watching, what happened there with my SMIRF engine, I had been able to recalibrate SMIRF's timing. Moreover I saw again an earlier known weakness during mating situations.
Now there is again a new bonus version of SMIRF MS-173 ready for download, which should behave a little bit better over all. Every owner of a personal SMIRF key set can download and use it from the well known address: http://www.10x8.net/down/SmirfEngine.dll .
Regards and thanks to HGM, Reinhard.
HGM's Winboard_F tournament
Moderators: hgm, Rebel, chrisw
-
- Posts: 484
- Joined: Mon Mar 13, 2006 11:08 am
- Location: Klein-Gerau, Germany
-
- Posts: 4556
- Joined: Tue Jul 03, 2007 4:30 am
Re: HGM's Winboard_F tournament
Thank you for the update!
-
- Posts: 27808
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: HGM's Winboard_F tournament
Well, I am happy that the tourney is useful! But you could also play such games yourself now. Last run was for testing if the new Smirf 1.72 still suffered time losses (it passed the test), but I will not always be playing Smirf.smrf wrote:Thank to Harm for his interesting automated tournament site!
Watching, what happened there with my SMIRF engine, I had been able to recalibrate SMIRF's timing. Moreover I saw again an earlier known weakness during mating situations.
Currently I am running a gauntlet for Joker80 playing at 4 min/game, against a variety of opponents with time advantages chosen to make them on average approximately equally strong as Joker80 (10 games per opponent). I will then repeat the gauntlet with another set of piece values, to see wich works better. The piece values I use now are simply those of normal 8x8 Chess, and I have the imprssion that this makes Joker overvalue Knights and Pawns quite badly.
After these Joker80 gauntlets, I guess I will be ready for the 'big event': the Battle of the Goths 1-hour games Championship 2008.
-
- Posts: 13447
- Joined: Wed Mar 08, 2006 9:02 pm
- Location: Dallas, Texas
- Full name: Matthew Hull
Re: HGM's Winboard_F tournament
To what extent is this automated? For instance, could you use chessiverse logic as a process wrapper for learning optimal piece weights?hgm wrote:Well, I am happy that the tourney is useful! But you could also play such games yourself now. Last run was for testing if the new Smirf 1.72 still suffered time losses (it passed the test), but I will not always be playing Smirf.smrf wrote:Thank to Harm for his interesting automated tournament site!
Watching, what happened there with my SMIRF engine, I had been able to recalibrate SMIRF's timing. Moreover I saw again an earlier known weakness during mating situations.
Currently I am running a gauntlet for Joker80 playing at 4 min/game, against a variety of opponents with time advantages chosen to make them on average approximately equally strong as Joker80 (10 games per opponent). I will then repeat the gauntlet with another set of piece values, to see wich works better. The piece values I use now are simply those of normal 8x8 Chess, and I have the imprssion that this makes Joker overvalue Knights and Pawns quite badly.
After these Joker80 gauntlets, I guess I will be ready for the 'big event': the Battle of the Goths 1-hour games Championship 2008.
Matthew Hull
-
- Posts: 27808
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: HGM's Winboard_F tournament
No, there is no feedback to the engine; I just use a standard tournament setup (using PSWBTM + WinBoard) to run round robins or gauntlets. I have to provide the engines / engine settings by hand.
I don't believe adjusting evaluation parameters and testing the effect of the adjustment is a competative way: most evaluation parameters have no effect at all in most games, and all those games contribute to the statistical noise.
So my approach is to directly measure the advantage (in terms of score percentage) conveyed by the various evaluation features (so far only piece values), by making sure they (and no others) are present in the initial position. For such tests I also have to provide the file with initial positions by hand.
I don't believe adjusting evaluation parameters and testing the effect of the adjustment is a competative way: most evaluation parameters have no effect at all in most games, and all those games contribute to the statistical noise.
So my approach is to directly measure the advantage (in terms of score percentage) conveyed by the various evaluation features (so far only piece values), by making sure they (and no others) are present in the initial position. For such tests I also have to provide the file with initial positions by hand.