Robotics Reference
In-Depth Information
to induce its opponents to make mistakes based on an incorrect model
of the program's play.
The original modelling system employed by Poki was based partly
on a set of numerical weights, one weight for each of the 1,326 pos-
sible combinations of hole cards. A particular set of weights indicated
the probability that the opponent would have played the current pot up
to the present moment, if he held this particular combination. These
weights were updated every time the opponent made a betting decision,
so the model for each opponent was always up to date, taking into ac-
count the opponent's play in the most recent hands. For example, a raise
increases the weights for the strongest hands likely to be held by the op-
ponent and decreases the weights for the weaker hands.
This method of modelling was found to be too simplistic because the
model did not take into account some of the most important details that
affect the decisions of strong players, such as the number of opponents
still in the pot, the size of the pot and the program's position in the bet-
ting order. An improved method also treated each opponent individually,
building one model per opponent. This made a significant difference to
the program's performance as it had hitherto treated all opponents in the
same manner, even though each player has their own style of play.
Because of the need for a successful Poker player to learn about his
opponents as the game progresses, the Alberta group investigated the use
of neural networks designed to predict the next betting action by each
opponent, based on a full history of that opponent's play during the
previous few hundred pots. The data on which the neural networks were
trained consisted of actual hands played by particular opponents. The
results of this investigation were that the actions of real opponents could
routinely be predicted with an accuracy of eighty percent and, in some
cases, as high as ninety percent
56
. This research also identified two highly
useful features for prediction that were added to the opponent model.
56
The accuracy was not quite as good when competing against very strong players.
