What I found most interesting was the part at the end when he was talking about the bias-variance tradeoff in using heuristics. It was kind of ironic - he was showing us a graph of the prediction error as a function of the order of the polynomial, and minimizing it. In other words, he was optimizing again. It's important to keep this in mind - heuristics aren't an alternative to optimization. They are a kind of optimization. What they are an alternative to is

*calculation*.

Speaking of heuristics, Michael Emmett Brady has reviewed the works of Daniel Kahneman and Amos Tversky before. He argues that while Kahneman and Tversky have their originality in experimental evidence, their heuristics are really a footnote to the decision theory laid out by John Maynard Keynes.

ReplyDeletehttp://www.amazon.com/review/RJA1JDCPMRN80/

http://www.amazon.com/review/RS1JM9M4FUF6A/

I'm not sure I'd characterize heuristics as a kind of optimization. I'd characterize it as a

ReplyDeletesolution methodbut that's not the same as optimization. Some heuristics use optimization, as you've pointed out (actually, to my mind, most do). But there are also heuristics that involve no optimization. For example: DO x IF condition y holds (concretely: buy an apple if it's sunny) is a heuristic that involves no optimization. It's just a rule of thumb.I'm talking specifically about how this speaker tried to convince his audience that using a heuristic was good - if you go to about two or three minutes before the end of the talk, you'll see he actually used an optimization technique.

DeleteRight, but that doesn't negate the overall point. And his example doesn't imply that all heuristics are a kind of optimization, which you seem to be saying. Again, a heuristic need not involve any optimization at all. If you think one of his examples was a poor choice in light of the point he was making, fine, but you seem to be saying something (and I think unjustified) more than that.

DeleteMoreover, I would say that

ReplyDeletealloptimization requires calculation. You can't optimize if you're not comparing alternatives, which necessitates calculation. That's where heuristics can differ - it may not require any comparison, and hence calculation (if I interpret what you mean by calculation correctly). It may just be a simple algorithm. In that sense, it isn't optimization.Throwing a forward pass is an optimization problem. Does it require calculation?

ReplyDelete