Here's a cool site: http://www.20q.net is a site that plays "20 questions." It has a learning engine behind it and gets trained by every user that plays it, so it keeps getting better and better... you hope. In truth, it probably won't keep getting better and better forever. First, its data is noisy, since some people just plain answer wrong (intentionally or not) and some questions are just vague (e.g. I was thinking of a quarter, and I answered that it was "mineral" -- it disagreed).
Second, learning systems tend to converge only so far, before more data just causes them to diverge again. I would suspect that is true here too. Supposedly 20Q uses a neural network; one of the classic approaches to stop neural networks from converging too early (or ever) is to add a little bit of noise by randomly tweaking a few of the weightings after each training session. Think of it as preventing the system from getting stuck at a "local maximum" instead of the "global maximum."
At any rate, it's a fun site. Check it out.
8:00:14 PM  ; ;
|
