AI Mastering 1-On-1 Poker

The powers of artificial intelligence seemingly know no bounds. Computers can make split-second decisions, use complex algorithms and use unlimited data to extract the exact information we need. They can assist everyone from writers to astrophysicists, and a recent poker experiment has revealed that they may even beat the world’s most skilled poker pros as well.

A team of researchers from the Czech Technical University at Charles University have paired up with the University of Alberta to create DeepStack, a piece of software that they believe can beat even the best poker players. So far, the research has focused entirely on no limit Texas Hold’em, but the software has performed exceptionally well thus far.

The researchers pitted DeepStack against 33 professional poker players who were singled out by the International Federation of Poker. Over 44,850 hands were played – and the computer won hands down. The game chosen was no limit Texas Hold’em, which has many more variations than its limited counterpart, showing the impressive skills of the AI software.

Researchers have even claimed to have solved Limit Hold’em using DeepStack technology. To further concentrate on Texas Hold’em, the team is in the process of writing software that will pit their computer against poker pros one on one. The complexity of playing against multiple opponents is still too complex for bots at this point, with the main win rate statistic used being mbb (mili-big blinds.) Poker typically involves antes, but in professional poker these are done away with.

In a casino setting, the dealer designates the big blind and small blind in each game, and pro games allow the designations of the blinds to change hands as the dealer does. Players often use the mbb/g statistic to indicate wins per game, and a win rate of 50 mbb/g is considered great, even by the professionals. The computer, in contrast, won an exceptional 492 mbb/g.

While these claims will surely excite any poker fan, they must still go through rigorous peer review testing before they can be considered legitimate. Many critics have already expressed scepticism, noting that the selected poker pros may not have been the world’s best, and that the timing of the announcement coincidentally came just days before a team from Carnegie Mellon University is set to use their AI software, Libratus, to play 120,000 hands against four poker professionals in a Pittsburgh casino competition.

In any case, one can certainly be impressed that DeepStack won against 33 top poker players over 44,852 hands, winning over 4 standard deviations above what could be considered a ‘good’ performance. The software’s success so far can only mean promising developments for AI casino algorithms in the future.

Play