[Bug-gnubg] Train bots for DMP, GG, and GS play ?

[Albert Silver]

Here is another idea posted in Gammonline by Neil Kazaross:

 

Posted By: Neil Kazaross <address@hidden>
Date: Saturday, 28 September 2002, at 7:31 a.m.

Here's some food for thought for our Bot developers here.

In my opinion I think some improvement in match play strengths and bot MWC can be had if bots are trained to play from the start both at DMP and GG/GS. This hopefully would aid to our knowledge of BG and improve human playing strength slightly as well.

Someone correct me if I'm wrong, but the way I understand that my 3 bots (SW,JF,GNU) select plays is that for each position they produce their table of WBG,WG,W,L,LG,LBG and then make the proper adjustments according to the values of G's and BG's won and lost at the match score. For example at true GG when gammon value is 1 and gammon cost is zero (BG's don't matter so I won't include in table) Let's suppose that play A and B are estimated by our bot to give the following:

A) WG .15 W.55 L.45 LG.10 B) WG .10 W.55 L.45 LG.05

It's very clear that for money and DMP the two plays are pick-em and for GG play A is correct and for GS play B is correct.

However, this evaluation table is produced by a bot which was trained playing cubeless for money and not at critical scores. My contention is that there's a trickle down effect throughout the game and learning process which might produce a better evaluation if the bot was trained at these critical scores.

So my first challange/question to the GNU developers is to train a separate GNU engine just for DMP ! Clearly playing two GNU's vs each other at DMP should result in 50.00 % wins for both sides after a long number of trials. My belief is that a GNU trained from the start to evaluate only at DMP will learn to do a better job of both evaluating and playing at DMP since there's never any noise from gammons/BG's and cubes to affect previous plays which could misslead the evaluation and learning process.

Maybe the DMP bot can win about 50.5 to 51% at DMP vs the "normal" bot. That would be significant and it might be making some instructively different plays.

However, if the difference is only about 50.1 % , it wouldn't seem worth the effort to include the DMP bot in the program.

If the DMP GNU bot proves sucessful, then we can move on to look at GG/GS.

I have a very fast PC and if the GNU developers decide to do a DMP bot, I'd be happy to use my 2.4 GHz for training it.

..ideas//comments// ??'s....neilkaz..