Tuesday, October 11, 2011

Stanford AI Class Unit 1w, 4 -- Terminology

I'm not a big fan of the AI Class style so far. The pen on paper animation I guess is a tried and proven teaching technique, but I was hoping that a course centered around software my use video t its full potential with animations, etc. But the basic "whiteboard" teaching approach is probably good for the target audience -- students that are used to learning in a classroom with chalkboard or white-board teaching. The first quiz was to classify the AI problem of playing a game of checkers in the 4 "dimensions" that Norvig seems emphasizes

    1. fully observable or partially observable
    2. continuous or discrete
    3. adversarial or benign
    4. deterministic or stochastic

Norvig makes the answers to this quiz clear in his video lecture, but there's room for introducing some subtleties, if Norvig had time. The mechanical action of actually playing chess brings in the oposite features to some degree. For instance, the AI agent should abstract the game of chess into a discrete form or state to facilitate processing on that state. But if the AI agent or robot is only given a camera as a sensor, then it would be discretizing or digitizing a continuous environment (image) and then atempting to discretize it further by identifying all the checker pieces and the boundaries of the squares and where they are located. Those pieces that are left halfway (either by the robot or the human opponent), might require some sophisticated AI to categorize. And if a new oponent from China or Japan started placing checker pieces on the intersections rather than the squares, well then, that could send the AI robot into a fit of confusion. You can think about how a physical game of chess might differ in 2 other categories as well. Can you see that there is a stochastic element to the physical moving of pieces? Can you see that a smoky room or a low aspect angle on the robot's camera could make the game partially observable? I can't think of a legitimate way to make the game of checkers benign, but the AI agent might have to deal with supportive training opponents that actually are attempting to help the AI agent win a few games to get it's learning engine going.

So it looks like this course is going to provide a lot of opportunity for side discussions and thinking -- if you have time on your hands like I do.

No comments:

Post a Comment