opening it up with Common Lisp
Book review: Darwinia
Summer reading: Spin
the Omnivoire's Delimma
the Golem's Eye
Principles and Applications of Continual Computation
That paper provides a pedestrian treatment of an uplifting idea: use excess computational power to help answer future questions. Although this isn't an original idea — I know of web browsers that attempted this at least 10-years ago and some of these were pretty sophisticated. — Horvitz does frame the problem in a formal enough way that he can prove things about how to minimize the total computational delay or how to maximize the quality of answers. These proofs fit into specific scenarios constraining the kinds of problems they can handle (e.g., are they all or nothing or do they fit into an anytime algorithm framework). He also takes into account memory, caching and the real-time nature of some domains.
Looking back over the paper now, I'm not completely certain what it is that I don't like about it. It is well written, complete, and moderately formal. Perhaps that is the problem: its treatment is too formal and sucks the life out of something that should be fun. I react to it in the same way I react to statistical treatments of language learning — yes, we need the formality and some parts of language learning are statistical but statistics aren't the whole story. All the talk of policies and optimization leaves me thinking that the focus is too syntactic. What about semantics and pragmatics? They are harder to talk about and perhaps impossible to formalize sufficiently but they are where our computations begin to matter.
This is a stretch (and a reference to only a so-so movie) but until we computer scientists and artificial intelligence researchers understand the anger of Will Smith's character in I, Robot, we're going to keep creating tools and applications that are brittle and that fail to meet the whole wonderful and wacky world we live in.
Copyright -- Gary Warren King, 2004 - 2006