Artificial intelligence researchers and computer scientists have been invited to take part in an unusual kind of guessing game.
Called the Machine Learning Challenge, the aim is to develop a computer program that beats all competitor programs at playing a game. The only catch is none of them are told the rules.
A private research company in Israel called Artificial Intelligence (Ai) has set the challenge. Its interest is to develop better ways of allowing computers to acquire language.
Advertisement
Key to this is being able to make flexible inferences about collections of words, without having to know all possible grammatical rules. Software that can learn the rules of games by inference alone could provide progress towards this.
Black box
In the competition, players are simply given a list of possible legal moves whenever it is their turn. They choose a move and then a score is given for how good that move was by an automated judge. 鈥淭he idea is to maximise your score,鈥� explains Jason Hutchens, Ai鈥檚 chief scientist.
鈥淚t鈥檚 a contest where people have to write programmes that conform to a 鈥榖lack-box鈥� interface just from observing symbols,鈥� says Hutchens. They are hoping to attract a wide range of approaches to the problem.
鈥淚 don鈥檛 hold any lofty expectations,鈥� says Hutchens. 鈥淚 think people will be surprised at just how difficult this is.鈥�
鈥淏ut I do think there鈥檚 going to be some interesting results, because I don鈥檛 think, for example, anyone鈥檚 played off a genetic algorithm against a neural network before.鈥�
Hidden strategy
Although, for obvious reasons, Ai will not discuss what kind of game the challenge will involve, Hutchens did say that it is likely to involve more than just one game. It could be as simple as tic-tac-toe, says Hutchens, but it is not likely to be as complex a game as chess.
People will also not know what their programmes are up against. 鈥淵ou could have an artificial intelligence graduate researcher at MIT competing against a 12-year-old hobbyist,鈥� says Hutchens.
If Ai decides to develop the winning programme commercially they will offer $25,000 as a sweetener to try to persuade the winning programmer to part with their game player.