Poker Bots Learn to Bluff

Poker-playing software programs are great at calculating odds and keeping, so to speak, a straight face. But the bluff -- that highest art of the game, the ability to intuit when and how to successfully play a low pair like a full house -- has always been beyond the grasp of their code.

"Computers are programmed to perform the best strategy, but bluffing is based on unexpected, illogical actions," says Evan Hurwitz, a computer scientist at the University of the Witwatersrand in South Africa.

All that may have changed: Hurwitz and his colleagues developed an artificial intelligence that learned to bluff.

Based on a neural network algorithm typically used to predict the stock market -- talk about unexpected and illogical actions! -- Hurwitz's bots weren't pre-programmed with the rules of a card game called lerpa.
(Yes, the bot doesn't quite play *poker*, but it's the principles that matter.)

Instead, they were pitted against each other and learned to play by inferring the game's rules from their own hands, those of their opponents and the outcome of the games. Eventually, one of the bots -- dubbed Randy -- suddenly started to bluff, having calculated that it increased his chances of winning against his still-cautious computer opponents.

"This demonstrates that computers can learn this peculiarly human behaviour," says Philippe de Wilde, a computer scientist at Heriot-Watt
University in Edinburgh, UK. "They generate the strategy from play, which is a very human way of learning."

So what else can a poker-playing, stock market-reading artificial intelligence learn to do? This definitely calls for Bruce Sterling.

Related Wired coverage here.

Software learns when it pays to deceive [New Scientist]

Image: Greg76