Re: current development

[Timothy Y. Chow]

On Sun, 8 Dec 2019, Joseph Heled wrote:
> Of course you need to weight every position with the probability it 
> occurs in actual play.

You say "of course," but I don't agree.  Weighting things in that way 
amounts to demanding perfection only from the starting position.  In my 
book, perfection means perfection from any legal position.  This is 
sometimes referred to as "strongly solving" a game as opposed to "weakly 
solving" or simply "solving" it.

> I don't think 2009 threads are a good indication. We need something with 
> the current net, which I think is better.

This is fair.  I would guess that GNU 2-ply (version 1.xx) and XG 3-ply 
(version 2.xx) are still susceptible to the tactic, though less so than 
earlier versions.  But this is just speculation; the only way to find out 
is for someone experienced with the relevant tactics to try it out.

I think that XG won't let you turn the cube past 1024 in actual play, so 
that might be an obstacle.  What typically happens in a money session is 
that the human loses a long string of games and then makes up for it in a 
favorable game, when the bot will beaver and redouble when it is losing. 
If you can get the bot to do this a few times in a row then you can win 
thousands of points in a single game.  If your goal is simply to come out 
ahead at the end of the session, then you might need to win just one such 
super-favorable game, since then you can protect your lead by dropping all 
doubles in all subsequent games, and refusing to double yourself until 
you're sure it's a drop.

Of course any computer is going to have *some* limit on the cube but I 
doubt that a cap of 2^30 or even 2^20 will be a serious limitation.

Tim