07-27-08 - 4

If you're an expert-system aggregator, you take in advice from various experts, weight them based on their confidence and your own prediction of expert accuracy based on local conditions. In practice it's best to have experts that have very strong opinions which are very different from each other. You don't actually really want the experts that have the best entropy on their own, since they will tend to be very similar and their opinions are already well hedged.

For example, in data compression, if you want to weight two models, you should pick something like one PPM and one LZ so they are very different. In Netflix if you are weighting two models, you want to pick one that's very local (similar users & movies matrix) and one that's very global (low rank SVD).

I see the same thing in human life decision making. I don't really want to talk to a smart advice giver who will see the pros and cons and give me a thoughtful hedged answer, like "well, there's this side and that one" - yes yes, I know that already. It helps a lot more to just be able to go to an expert who presents one side of the argument very strongly. You can listen to that and consider it. Then you go listen to an expert who presents the other side. Then you make your own decision.

No comments:

old rants