Computers have beaten human world champions at chess and, earlier this year, the board game Go. So far, though, they have struggled at the card table. So we challenged one AI to a game.
Why is poker(扑克)so difficult? Chess and Go are “information complete” games where all players can see all the relevant information. In poker, other players' cards are hidden, making it an “information incomplete” game. Players have to guess opponents' hands from their actions—-tricky for computers. Poker has become a new benchmark for AI research. Solving poker could lead to many breakthroughs, from cyber security to driverless cars.
Scientists believe it is only a matter of time before AI once again vanquishes humans, hence our human-machine match comes up in a game of Texas Hold's Em Limit Poker. The AI was developed by Johannes Heinrich, researcher studying machine learning at UCL. It combines two techniques: neural(神经的)networks and reinforcement learning(强化学习).
Neural networks, to some degree, copy the structure of human brains: their processors are highly interconnected and work at the same time to solve problems. They are good at spotting patterns in huge amounts of data. Reinforcement learning is when a machine, given a task, carries it out, learning from mistakes it makes. In this case, it means playing poker against itself billions of times to get better.
Mr Heinrich told Sky News: “Today we are presenting a new procedure that has learned in a different way, more similar to how humans learn. In particular, it is able to learn abstract patterns, represented by its neural network, which allow it to deal with new and unseen situations.”
After two hours of quite defensive play, from the computer at least, we called it a draw.
