Poker playing artificial intelligence has already "approached the performance" of human experts and can use "state-of-the-art methods" in its gameplay.
Researchers from University College London - including a staff member from DeepMind's Go defeating team - have created a series of reinforcement algorithms that are able to play Texas Hold’em and a simplistic Leduc poker.
The AI is able to learn the game without any prior knowledge of strategies and taught itself by playing fictitious matches on its own, according to the paper Deep Reinforcement Learning from Self-Play in Imperfect-Information Games.
Research student Johannes Heinrich and lecturer and David Silver explain in the paper that the Neural Fictitious Self-Play method they created used deep reinforcement learning "to learn directly from their experience of interacting in the game". The method learnt from its mistakes and developed ways to win the games, while also utilising neural networks.
The researchers claimed their model was able to simulate the Nash equilibrium for Leduc, while a similar breakthrough was close for Texas Hold'em. "It [is] conceivable that it is also applicable to other real-world problems that are strategic in nature," Heinrich told The Guardian.
The research paper comes after Google's DeepMind AI defeated Go world champion Lee Sedol 4-1. The AlphaGo AI managed to beat Lee by playing moves that humans were unlikely to make or be able to predict. Despite the defeat Lee has since said he would take on the deep learning system for a second time.
While Go has been conquered, poker presents various other challenges for those developing AI. The unpredictability of humans, being one.
The UCL researchers aren't the first to try and create general learning algorithms to defeat poker. In 2015 a 14 day, 80,000 hand competition of Texas Hold'em saw AI from Carnegie Mellon University take on humans for the first time.
In the contest humans came on top by $732,713 after a theoretic $170 million was wagered by both sides. The AI was hindered by how it responded to humans raising the stakes; the unpredictability of human betting meant that the AI was having difficulties interpreting the games.
Humans also were able to take advantage of the AI's inability to predict why the cards in a person's hand may impact the game. As such it was easy for the humans to tell when the computer program was bluffing on a weak hand.
This article was originally published by WIRED UK
