Computer AI Masters Poker

An artificial intelligence program has beaten professional poker players for the first time. It may sound trivial but marks a major advance in technology that could have practical benefits.

Computers beat human experts at chess many years ago, and last year a computer beat the world champion at the far more complex board game "Go" for the first time. That was significant because the sheer number of possible moves available in Go means ever the most powerful computer can't assess every possible move.

Instead the program had to learn to use human tactics such as heuristics (in effect, rules of thumb) and trying to find out ways to rule out large sets of options without having to consider them one by one.

On the other hand, Poker posed a very different challenge. The computer program, "Libratus," had to deal with the fact that in the variant played in the experiment - heads-up no-limit Texas Hold'em - it faces a huge number of different combinations of information - namely, the cards it holds, plus any cards revealed on the table. In fact, the number of possibilities is 10 followed by 160 zeroes.

There's also the issue that whereas chess and Go are perfect information games (where both players know where all the pieces are), poker is all about imperfect information, meaning players doesn't know what cards their opponents hold.

Computers Learn Art of Bluffing

To add to the complexity, the decisions in poker aren't about improving a hand of cards, but rather on how to bet on them, with bluffing being the key. Libratus not only had to learn patterns to try to decipher the likelihood that an opposing player was bluffing, but also to work out when to bluff. Unlike human players, it couldn't detect physical "tells", the unconscious physical signs some players give when they are in a strong position or are bluffing.

The advantage is that a computer can play for long periods without losing concentration, and that it doesn't make emotional decisions such as letting fear of losing a large pile of potential cash winnings affect its judgment.

Business and Military Could Benefit

After a defeat last year, Libtratus beat four human players in a game that lasted 20 days, a duration designed to minimize the effects of luck. (Source.: bbc.co.uk)

Computer science experts said the result could mean it's more viable to harness the benefits of artificial intelligence in areas with imperfect information. For example, it could be used for negotiating deals or prices in situations where each side doesn't know how much the other side would be willing to settle for and how likely it is to walk away if it is not possible to get a good deal. Artificial intelligence might also be useful in decision making in the military where one side only has partial knowledge of the battlefield. (Source: zdnet.com)

What's Your Opinion?

Are you surprised computers can now perform well at poker? Is such game play a sensible subject for research? Do you welcome or fear computers being used for real-world decision making?