next up previous contents
Next: Thesis contributions Up: Introduction Previous: Why games?

Why poker?

So far, the primary focus of games researchers has been placed on algorithms to solve games with perfect information. As a result, high-performance systems have been developed for games such as chess, Othello, and checkers. In many of these games, high performance can be achieved by brute-force search. Recently, attention has been given to games with imperfect information, such as bridge and poker, where searching seems not to be the key to success. Since these games offer different algorithmic and conceptual challenges, the successful development of a program capable of playing them well may provide solutions to open problems in computer science.

Poker has several features that make it attractive for AI research. These include imperfect information, multiple competing agents, risk management, opponent modeling, deception, and dealing with unreliable information. These characteristics are also present in many real-world applications that require rational behavior.

Imperfect information implies that a choice must be made from a set of actions without complete knowledge. The relative desirability of each action depends on the state of the world, but the agent does not know exactly which state prevails. In poker, a player does not know the opponents' cards. Without knowing the complete state of the world, how can the player find which actions are, ``optimal'', in some sense?
Having multiple competing agents exponentially increases the complexity of the computations required to play poker by enlarging the game tree.
Risk management requires making a decision to gain a profit while considering how much one can afford to lose. Making a good decision based on the evidence available and ``cost-benefit'' considerations is a skill required in many real-world activities. For instance, investing in the stock market has the same adrenaline-releasing characteristic. Every time a player makes a betting decision in a poker game, there is the risk of losing money. However, there is always a chance to win. In the long run, a player's objective is to end up with a positive balance.
Opponent modeling involves identifying patterns in the opponents' play and exploiting any weaknesses in their strategy. For example, opponent modeling is extensively applied in political campaigns. In poker, it can be done by observing the opponents' betting habits, and determining likely probability distributions for their cards. If a player can predict the opponents' actions, then this player will be capable of making much better decisions.
Deception and the ability to deal with unreliable information are traits of a strong poker player. In fact, these activities are also necessary in real-world situations. For example, assume one wants to acquire a used car. How much shall one believe from all the wonders the salesman says about the car? How can one get a reduction on the price of the car? Good poker players have to be unpredictable by bluffing and varying their playing style, and must also be able to deal with their opponents' deceptive plays. For example, if a player is known to raise only with a strong hand (a predictable player), the opponents are likely to fold in such cases. Therefore, this player is missing opportunities to earn more money on the best hands. By occasionally raising on a weak hand, this player will either profit from a successful bluff, or will implant doubt that will result in greater profits for strong hands. Hence, it is necessary to mislead the opponents by letting them know that an occasional raise or high bet is possible with a weak hand.

next up previous contents
Next: Thesis contributions Up: Introduction Previous: Why games?
Lourdes Pena