I just found the post explaining how this was accomplished (see
here):
[ QUOTE ]
Some questions about Sparbot were raised in another forum,
regarding the aggressiveness settings, and about occasional
bizarre plays by the bot. Here is the reply I posted.
...<mike@a...> wrote:
>
> Sparbot has an aggression bar you can set.
>
> Am I to assume that no matter how you set this aggression bar
> (e.g. either most passive setting, or most aggressive setting),
> it's "objective EV" will be the same (or very very close)?
>
> something tells me the answer is no, because:
>
> I ran a ~25,500 hand simulation of Sparbot (i.e. the bot at the most
> passive setting) vs. Sparbot2 (its aggression bar is approximately
> 66% of the most aggressive you can set it to). The results were
> alarming. Sparbot ended up winning 0.102 sb/hand! Not only that,
> it achieved +EV even when it was in the BB (i.e. out of position)!
>
> Anyone care to explain? Aaron? Darse?
When the linear programming (LP) solutions were computed,
they had the same objective EV. However, the resulting strategy
is only a crude approximation of an equilibrium strategy, and
a good player can find serious flaws and exploit them.
The aggressive solution takes more risks, and is thus more
vulnerable to being soundly beaten. It turns out that when
they play against each other, the more passive Sparbot (the
original version) is well-suited to exploiting the errors of its
cousin. However, the more aggressive version is much more
appropriate against the vast majority of human opponents.
The difference isn't as large as your results suggested, but
yes, one bot does beat the other. One data point against
one particular opponent is neither here nor there.
There are also small residual probabilities of making some
bizarre actions, due to numerical stability issues in the LP
solution. The mixed strategies in certain situations might
have a 0.001% chance of calling with an extremely weak
hand, or of folding a very strong one. The more aggressive
solution had more of these problems than the original, but
we never took the time to clean up the solutions after the
fact, since it has only a slight effect on the bottom line.
[/ QUOTE ]
[ QUOTE ]
Sparbot2 was solved to have the same objective EV as Sparbot1 in the abstract game that we use as a model for real poker (with the additional constraints that it is more aggressive). Since there is a mapping procedure from the abstract game to the real game of Hold'em it is hard to know which one is closer to optimal in terms of the real game (the game that they actually play and that you play against them).
[/ QUOTE ]
Juk [img]/images/graemlins/smile.gif[/img]