Re: [Bug-gnubg] The little engine that could... Neural Network

[Philippe Michel]

Øystein Schønning-Johansen  wrote:

> What about other techniques? [...] Simplified evaluation function 
> combined with deeper search?

I don't think this is very promising. With 21 rolls at each ply and very 
few forced moves, deeper search is very expensive.

Moreover, gnubg uses very radical forward pruning, with only the best 
statically evaluated move expanded at each ply. No minimax/alpha-beta or 
their probabilistic variants at all, so the static evaluation has to be 
pretty good.

In fact, maybe we could try to explore the game tree a bit more. For 
instance, if the best move at an intermediate ply is a hit, check the best 
non-hit as well (if it doesn't look a lot worse) and vice-versa.


Regarding the evaluation function, we could review the inputs. Besides the 
raw inputs, they are mostly (only ?) 40 year old ideas from Berliner. 
Joseph Heled apparently tried some things but found it hard to improve on 
the existant. XG apparently has 256 inputs (vs. 250 for gnubg) so it must 
use more and/or different inputs. In an old article in Inside Backgammon, 
Brian Sheppard wrote "We figure a typical net has 400 inputs" but this 
looks like a random guess. I think he is the author of a very strong 
Scrabble program and was knowledgeable on games AI in general, but maybe 
not specifically neural networks and backgammon.

A side issue with the inputs is that a few of them are extremely expensive 
to compute. Maybe they could be replaced by simpler ones (possibly replace 
one complex input by a few simpler ones and let the network emulate the 
former).