Re: "20q" in perl?

[eric256]

I can't find any confirmation but I'm pretty cure 20q.net is NOT a nueral net. There is no need for it to be. Basicaly it breaks down to this. You have a set of objects and a set of questions. For each object you have 1 or more questions (more = better) that you know the answers to. Start with a random question (hopefully one that divides the rest of your items in half). Then continue to ask questions that divide the rest of your items in half (half say yes and half say no). You should reach the bottum of your tree and be left with the item that the user is guessing. There are some finer points since some items wont have answers to all questions. Obviously though the more questions and items/answers you have the smarter your system seems. I've programmed one of these in the past and they are quite fun and seem quite intellegent.

[schweini]

i was just guessing "neural" because i just assumed that 20q does handle some kind of 'link strength' between questions and answers so that a question that gets the same answer 20 times makes that answer more 'important' than some random answer given by a lost user. and IIRC, that was the basic mechanism of neural networks - that repeated 'stimuli' reinforce a 'link' - although i might be completly wrong (i still think it would make sense, though: wikipedia works because they assume that most users will actually do the 'good' thing, and the minority that do nonsense just dissapear under the flood of good stimuli/edits)

i somehow assumed q20 was open source or something, but i guess i was wrong. :(

one possible application i saw was for medicine, because looking over my gf's shoulder while she was studying, i noticed that a lot of diagnoses seem to be kind of fuzzy, and have a lot of interconnections with other symptons. (e.g. high blood pressure with diabetes is different from high blood pressure with only smoking), so i thought an 'intelligent' q20 thing would at least be a help for doctors.
on the other hand, i once heard that medical expert systems failed because too many diagnoses are actually made intuitivly/subconciously, due to the sheer amount of possible sympton-diagnose relationships.
ah well. would still be interesting, i guess - anybody know of some machine-readable symptom -> diagnosis graph?