next up previous contents
Next: Action Selection Up: Betting Strategy Previous: Betting Strategy

Bluffing

A minimal bluffing technique would indiscriminately bluff a small percentage of games. The original implementation of BPP made use of this simple technique to limited success, with one obvious flaw. It would also indiscriminately choose to bluff within a single game. Occasionally choosing to bluff in the final betting round, and within the same round, choose to not bluff with its following action. This flaw in play has been eradicated by introducing a bluffing state. BPP now has a low probability of entering a bluffing ``state of mind'' in the final round of betting. Once in this state, BPP will continue to bluff until the end of the round (and hence game) making for a more convincing bluffing technique.

But this method still had many weaknesses. Occasionally BPP would bluff when holding an obviously weaker hand than its opponent, with potentially no way to improve the strength of its hand above that of its opponent. An example is when its opponent might be revealing a 3-of-a-kind in the last round of betting while BPP has a busted queen revealed. In this instance BPP should not attempt to bluff as it is highly unlikely a good opponent will fall for the bluff.

An improved bluffing technique would not only bluff a small percentage of hands, but also consider such things as hand potential and the likelihood that the opponent will fall for the bluff. This has been addressed by comparing the visible hand of both players in the final round and their potential. Using the Bayesian belief network, evidence is added about both players visible cards. The belief in winning the game at a showdown is then observed and used to decide whether or not a bluff could potentially succeed. If the belief in winning is zero (there is potentially no way for BPP's hand to improve above that of its opponent) bluffing is not considered. If the belief in winning is greater than zero, BPPs hand could potentially defeat its opponents hand (by observing only the visible cards of both players) and therefore has some chance of a successful bluff and will bluff with a low frequency.

To facilitate this technique, it was necessary to add an additional node to the current network structure. This node, BPP Upcards, represents the visible hand of BPP and was added as a child of BPP Current so it may be used in inferring the potential strength of its hand (from the point of view of it's opponent) when considering a bluff.

The addition of an improved bluffing technique was not based an attempt to achieve immediate performance gains, but was rather based upon critique by experienced human players. The bluffing decisions made based upon the original previous strategy were determined to be ``unreasonable'' after review, and modifications were made were eradicate the serious flaws in play. Due to the small number of games in which a bluff is made, there is no expectation that it will improve play significantly in the short term (particularly in self play experiments) but rather has the purpose of creating doubt in the mind of it's opponents and represents a small part of the total package required for successful poker play.


next up previous contents
Next: Action Selection Up: Betting Strategy Previous: Betting Strategy
Jason R Carlton
2000-11-13