🚀 go-pugleaf

In article <5dsm20$s4s@ausnews.austin.ibm.com>,
Jim Babka <babka@austin.ibm.com> wrote:
>Actually, I have a theory about this.  I'm not "in the know" about this - I
>don't know the people who put TD-GAMMON together - but I have noticed the
>rather blatant appearance of cheating (program has three unprotected pieces,
>only way I *won't* hit it at least once is if I roll a 1 2, and guess what I
>roll).  My theory is that this is a game that was trained via massive
>reiteration of playing against itself.  During the course of that learning,
>it also "learned" the pattern to the pseudo-random numbers used to generate
>the die rolls.  I understand that they probably change the seed from game to
>game, but depending upon the quality of the random-number generator, there can
>still be a repeating pattern, and that pattern might be built into the neural
>net (or whatever other algorithm is being used) that drives the computer play.
>

a) you're right about the technique used for building TD-gammon (last time I 
   heard, it was a self-trained neural network.

b) you're almost certainly wrong in your suggestion that TD-gammon acts based 
   upon learned knowledge about the way the dice perform.

  What you're proposing (as I understand) it is that TD-gammon gives high 
equity ratings to positions which, though bad, are made good by a persistent 
bias of the dice, and thus are 'safe'.  For this sort of persistent bias to 
exist, it would appear to be necessary for the dice to be linked in some way to
the positions being seen, a near-impossibility.  Given that TD's evaluation of 
positions is based upon a 'generalization', if you will, of a plethora of 
somewhat similar situations, I rather suspect that you could get away with 
using a decidedly _non_ random die sequence for training, and that you would 
still end up with a fairly valid evaluation engine, i.e. one which is not 
biased towards certain risky situations.

  As further evidence that no such bias exists, witness the fine performance 
of neural net-based 'bots like td-one, jellyfish, and mloner on FIBS, where the
die-rolling algorithm is almost certainly not the one that these 'bots were 
trained with.

  How about this: I'll give you ten to one that you're mistaken.

john