bug-gnubg
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] Computer Olympiad


From: Frank Berger
Subject: Re: [Bug-gnubg] Computer Olympiad
Date: Thu, 21 Jun 2007 19:41:30 +0200

Hi Øystein,

Yes!

I think we should view the Computer Olympiad as a tournament (as it's
actually is), and not as an analysis of which program is best.
Total agreement agree.



Do I read between the lines that you have matched the two against each
other? What's the score?
S4, JF, GnuBG 0.15, BGB  1000 25 pt matches, each against each.

As I havn't done it but just has informed the author should publish his results. I guess it will be in the next few days.




Frank? Can you give some details about your training algorithms. Are you
using TD-training exclusively?

Sure. I use plain TD-Lambda. SOmetimes Lambda=0, sometimes lambda = 0.1 or 0.05 switching to 0.0 later. Different to what was said here from time to time it works really amazing. The downsides: - one tiny error and it doesn't learn at all or it stays on intermediate level. As an example: I had improved aigammon (another AI derived from eggammon) and it was stuck at the level of the original ai..... Many month later (when I had already developped TachiAI) I found the bug accidently...

- sometimes it doesn't learn at all. It happens rarely.

I tried to train it with additional positions and with starting from Nackgammon. If you train long enough Nack- and Back- seems to converge (weired but that was the result) and the net with 1500 starting positions had no improvement from the net starting from the normal position, when playing against the normal net. There are different moves, but not a real improvement, just changes. Those experiments were made some years ago, so with the pc of that time the tests couldn't made that long, so it has statistical good significance.


I really appreciate that
someone takes the time to do this kind of thing. Even though it seems
that we do more to promote BGBlitz than anything else.
Well, I would see it the way that we promote Backgammon or Computer Backgammon. The promotional effect of the 2nd and 3rd win is *very* remote to the normal world and the BG savy people are to a high percentage in the GnuBG or Snowie camp anyway... altough I keep working on it ;)


ciao
Frank





reply via email to

[Prev in Thread] Current Thread [Next in Thread]