I think it would be good to look at ways of improving 0-ply cube evals.
Here are some ideas posted on the Gammonline Forum"
To measure volatility without lookahead, I can think of various ways. Very roughly is to just look at the breakdown, and adapt volatility from some standard figure by means of the gammon rates f.i. This might work just a little I guess; think of a blitz double with relatively low winning chances but high gammon chances. The bot could determine from the breakdown that volatility is probably high, and double earlier accordingly. More sophisticated is to use characteristics from the position itself. In fact this is not really different from what a NN does to determine equity with no lookahead, so one could train a NN to estimate volatility too. Without a NN trained for volatility available, one could use hand-crafted rules but that might be a bit tricky I suppose. Think of some very simple rules like: lower pipcount(s) is higher volatility; more blots is higher volatility; more innerboard points is more volatility; more men back is less volatility, etc.
Snowie doesn't appear to have any way of determining this but good old JellyFish did.
It was able to distinguish position types well at 0-ply and double accordingly if the volatility was high.
It always showed the (different) volatilities if asked at 0 ,1 and 2 ply.
Does anyone know how this was coded so hopefully something similar or better
could be used in GNU to complement it strong 0-ply play.
Michael
[Prev in Thread]
Current Thread
[Next in Thread]
[Bug-gnubg] Improving cube evals at 0-ply,
Michaeldepreli<=