Re: Programs that learn their opponent's weaknesses

On Feb 6, 4:03 pm, tc...@xxxxxxxxxxxxx wrote:
In article <af73c19b-8527-43da-b22e-82fc94850...@xxxxxxxxxxxxxxxxxxxxxxxxxxxxx>,

 <pauldepst...@xxxxxxx> wrote:
But this has got absolutely nothing to do with "experimenting" or "testing".

O.K., poor choice of words on my part.  I think we now understand each other.

It seems to me that if the best backgammon bots are based primarily on
neural nets and other learning-theory ideas, rather than on branching
and pruning, it shouldn't be excessively hard to incorporate some Bayesian
model of the opponent's play.  But what do I know?
--
Tim Chow       tchow-at-alum-dot-mit-dot-edu
The range of our projectiles---even ... the artillery---however great, will
never exceed four of those miles of which as many thousand separate us from
the center of the earth.  ---Galileo, Dialogues Concerning Two New Sciences

It just seems to be that no one involved in programming backgammon has
bothered to look at modelling the opponent's play.
There are trivial improvements that can be made in that area but no
one is working on it. For example, one could look at the average
equity loss the opponent has from dropping good takes and compare it
to the average equity loss the opponent has incorrect takes.

If there are many more bad drops than bad takes, the bot should double
even if doubling loses .05 in equity. And in later modifications, ".
05" could be replaced by a variable etc. etc.

Since, currently, there is _no_ programming of this type, it shouldn't
be hard to improve on it.

I think the reason there is no such programming is that consumers want
their bots to be theoretically correct, and to eschew psychology etc.

In chess, it is also trivial for a bot to exploit the opponents'
weaknesses. For example, it can play the openings that the human
opponent plays badly. I think we see exactly the same situation in
backgammon -- that no chess programs do this. (But I may be wrong.)

Paul Epstein


.

Relevant Pages

  • Re: Hand Discussion
    ... Why gamble all our chips away when we can play poker ... How many times do you see an opponent flop top pair and continue to fire ... ones who made a mistake because we didn't make him pay for his draw. ...
    (rec.gambling.poker)
  • Re: Hand Discussion
    ... what if he was drawing then hit? ... Why gamble all our chips away when we can play poker ... How many times do you see an opponent flop top pair and continue to fire ... ones who made a mistake because we didn't make him pay for his draw. ...
    (rec.gambling.poker)
  • Re: Situational superko vs. Positional superko
    ... in that case the opponent is apparently not playing ... > One can have two hypotheses about the play of the opponent. ... > can use ko-threats is readily explained if the second hypothesis holds, ...
    (rec.games.go)
  • Re: Playing to an opponents weaknesses
    ... concept is a sound one if you really know your opponent. ... In the Monty vs Mpetch series Monty has suggested he is the superior ... Over the series of 7 matches I ended up with a lower checker play error rate ... With Monty at 1950 and myself at 1900 he would be the better player. ...
    (rec.games.backgammon)
  • Re: What if bot analysis was taken to the next level?
    ... Making an intentional error in order to get your opponent to make a bigger error is one way to gain equity on your opponent...... ... So in order to do this in real life you either have to know your opponent so well that you know exactly what he'll do or see some clear concrete tendencies that you're quite sure he'll repeat. ... Where you can maybe take advantage of someone is if you notice certain things, like a somewhat strong player passed a very big take in a specific type of game. ... So if you notice he passed a big take from a certain type of position, you may want to try to play your checkers to increase your chances of getting back into that type of game to see if you can take advantage of his tendency even more efficiently. ...
    (rec.games.backgammon)
Loading