Tuesday, October 11, 2011

Stanford AI Class 1w, 10 -- AI as Uncertainty Management

Sources of uncertainty that AI can help overcome to produce effective decisions:

    1. sensor limits (Norvig means "limitations" not limits)
        a. resolution
        b. precision or measurement error (random noise, deterministic errors, cross coupling with other observables like temperature)
        c. dynamic range or range limits (min and max measurable values)
    2. adversaries (you can never be 100% certain what your oponent might do)
    3. stochastic problems or enironments
    4. laziness
    5. ignorance (people and their AI agents don't attempt to know everything)

Sensor limits (#1 above) are *not* a source of uncertainty. I'm a controls engineer, and we deal with sensor and actuator limits all the time with non-AI systems, as they are totally predictable and relatively easy to compensate for.

Adversaries (#2 above) are only a source of uncertainty if they act stochastically or even deterministically but chaotically -- where there internal decision system is so complex that you couldn't predict their future actions even with perfect and infinite recordings of their environment and past history of actions. Read more about Chaos if your are curious. The classic example is that even if you had a perfect and infinite supercomputer capable of modeling all of the "billiard ball" molecules of air throughout the globe, a small perturbation of any portion of those billiard balls (like a butterfly flapping it's wings in Africa) can have dramatic and unpredicatable affects on the weateher around the world (say a hurricane in North America). Nonetheless, Norvig is ignoring some extremely logical and predictable adversaries that can be treated as entirely deterministic and they will not add any uncertainty to the AI problem. The field of mathematics called "Game Theory" deals with this category of adversarial AI problems nicely, providing "closed form" solutions for the best actions (most rational action) for an AI system competing in an adversarial but deterministic world.

Stochastic environments (#3) are clearly a source of uncertainty, but I don't quite see how "ignorance" fits in (#5). I guess we'll have to gain some of Norvig's vast experience before we can appreciate what he means here.

No comments:

Post a Comment