[Bug-gnubg] Training Bots At Match Scores

[Øystein O Johansen]

Hi,

Michael asks this question to the GammOnLine discussion forum:

-----8<-----
Am I right in thinking all neural net bots are trained at cubeless money
play?



Has any one ever tested training a bot at say DMP?





Would it be noticeably better?


-----8<-----


(I'm transfering the question to this mailing list, and I really hope this
discussion will become interesting!)





Yes, you're right. Bots are trained at cubeless money play. At least the
bots I know something about.





I'm not sure if anyone have tried training a bot at DMP. Maybe some other
bot developer has, but I don't thing Joseph or Gary has trained any bots at
DMP. (correct me if I'm wrong). There is so many bots around, so I guess
someone must have tried. The idea is not new. Douglas (I think it was him)
suggest four types of nets. DMP, GS, GG and the ordenary nets.





Would it be noticeable better? I don't know, but I like to believe it's
going to be better. Lots of things are very different at DMP. It's often
good to run from an anchor, even though you risk beeing attacked. A
backgame is often more powerful, since you can keep the contact to the end,
without risking to loose a gammon or backgammon.





I would love to see how a special trained DMP network performs against a
traditional net. Will it be much hacking into the GNU code to train a DMP
net? NUM_OUTPUTS set to 1. Will it be much more?





-Øystein




-------------------------------------------------------------------
The information contained in this message may be CONFIDENTIAL and is
intended for the addressee only. Any unauthorised use, dissemination of the
information or copying of this message is prohibited. If you are not the
addressee, please notify the sender immediately by return e-mail and delete
this message.
Thank you.