Slashdot Mirror
Machine Learns Games
heptapod writes "New Scientist is reporting that UK researchers have created a computer that can learn rock, paper, scissors by observing humans. CogVis uses visual information to recognize events and objects in addition to learning by observing."
You sure you RTFA? The computer doesn't learn how to play, it just learns how to determine who won. That's not very impressive at all, considering the game was played with cards instead of hands, and there are only 9 possible hands and three possible outcomes (Left wins, Right wins, or Draw).
So the computer sees "Scissors-Paper" a few times and then always queus up the "Left Wins" response when it sees "Scissors-Paper" in the future. That's just a different method of programming.
Now, if only 6 of the 9 possible hands had been played, and then a 7th hand the computer hadn't seen before was played and the computer could tell you who won that, that'd be something. This is just record and playback.
paintball
It's not like they got a camera, gave it AI, pointed it at a rock-paper-scissors game and commanded it to "learn."
Granted, the parent poster is being silly, but that's actually not too far from what they did. They basically took the system and pointed it towards the people playing the game without telling it explicitly what to expect. From the article:
Chris Needham, another member of the CogVis team, says the system's visual processor analyses the action by separating periods of movement and inactivity and then extracting features based on colour and texture. Combining this with audio input, the system develops hypotheses about the game's rules using an approach known as inductive logic programming.
"It was very impressive," says Max Bramer, a researcher at Portsmouth University, UK, and chair of the British Computer Society's AI group. He told New Scientist that CogVis could have many future applications. "You can think of lots of times when you'd like to be able to point a camera at something and have a computer interpret things for itself."
He suggests that machine's could one day use this technique to learn how to spot an intruder on video footage or how to control a robot for important maintenance work. "It's a very good start, and almost mysterious in the way it works," Bramer adds.
From their page:
In this piece of work we are attempting to learn descriptions of objects and events in an entirely autonomous way. Our aim is zero human interference in the learning process, and only to use non scene specific prior information. The resulting models (object and protocol) are used to drive a synthetic agent that can interact in the real world.
Since primitive machines were invented, they always had a nasty habit of choosing A, B, human instead of A,B,C. I guess you didn't give much thought to human fingers in hot dogs or robotics-related industrial accidents in Japan.
The problem is precisely the lack of free will and independent thinking. A machine has grappling hooks, vacuum suction or serving belt, but it can not make value judgment on what/whom it is throwing into molten metal.
As the AI develops, the problem will get worse before it gets better. A robot working in slaughterhouse might have the ability to chase a running mammal and cut it's throat, but not to ascertain exact species. Imagine a beowulf cluster of those on the run in New York subway. Workspace and consumer safety legislation would be very much in order at that point.
If you actually want to understand what they did, some research publications put out by the CogVis lab have better information regarding the technical side of things.
Towards an Architecture for Cognitive Vision Using Qualitative Spatio-temporal Representations and Abduction (Cohn et al, 2003)
Modeling interaction using learnt qualitative spatio-temporal relations and variable length Markov models (Galata et al, 2002)
Watch the bugger doing it - I got knocked back for an internship by these dudes, but I did get to see the system.
It's bloody amazing, the amazing bit being it deduces how to play from first principles, starting with just the ability to identify that what it's being shown is an object.
Takes about 30 minutes to get rolling, but it really is stunning to watch! Hell, object differentiation is hard enough, deducing the rules of play, and tactics as well?
fortune -o
Well, it could just watch several people *pretending* to be intruders.
It can learn games other than RPS. The point is that you give no information about what you're playing at all - just give it the video and the other inputs coming from the players and it tries to work out what's going on from scratch.
When presented with a new game, the first thing you do is learn how to play. The second thing is to learn the best strategy to win. This is just concerned with the first part and that's the novelty - looking at human actions and learning what they're doing, not necessarily how to beat them.