i downloaded an app that’s similar to the handheld 20 questions game i had as a kid.
i’m thinking of the most obscure characters and it still guesses it every time. someone please explain to me how this works
In: Technology
They learn. There are millions of people playing the game right now. Occasionally someone does manage to think of a character that the app can’t guess. When that happens, the app remembers the person that picked, and the answers they gave, and now it’s ready for the next time someone picks that character.
It is an algorithm. Each question mathematically eliminates millions of answers until the possible answers are shaved down to one (or just a few)
A person thinks of any Thing. Any noun. No proper nouns. Othe other person or people get to ask a total of 20 yes or no questions.
Is it a mineral? no
Is it an animal? yes
Is it cold blooded? no
Is it a mammal? yes
Is it normally more than 50 lbs? no
Is it normally more than 5 lbs? yes
Is it a common pet? yes
Is it a dog? no
Is it a cat? Yes You win.
It’s been mentioned that the games have a big database of characters already. There are usually questions to filter out huge numbers of possibilities all at once, too determine if you’re trying to use a common character or an obscure one. Pay attention to the questions and you may start to get an idea of the scope of possibilities each one is trying to eliminate with a single question.
If you can knock out real or fictional, then you can narrow it to genre, then focus on questions that are hard to evade.
I actually learned this recently. The original 20Q device was actually built on an artificial neural network created and trained on a website that did the same thing the little electric game did. After training it with real people, using real rounds of questions it got really really good at figuring out what a person was thinking of. The network was trained in a way that didn’t rely on having everything programmed in. It actually learned to interpret how the user thought about questions. The more you used it, the better it got. It could even figure out what you might be thinking of based on more ‘folk taxonomy’ or how we as users might describe things, not just their standard or official clarification, i.e. tomatoes are technically fruit but it figured out most people would label them vegetables.
On a more speculative level I think the handheld game also had a neat advantage that helped it in ways that aren’t immediately apparent. The first is the massive library that it could store, without needing much more than general categories. An animal? Yes. That stops it from considering anything without the internal tag of ‘animal.’ The second is that it doesn’t need to ‘understand’ what it’s thinking about. It’s not working in the same direction as humans. We have to get an idea together, then determine questions that include or exclude it. But the computer simply has to take into account the number of possible answers it knows, and what questions best divides that group into efficient smaller parts. The third thing is something I specifically remember about the one I had growing up, but didn’t think about until much later. Mine, and I assume most of them, don’t have a yes or no as the first question. You could pan through animal, mineral, etc. Then you select one. So it gets a couple questions in.
If you can ask a question that divides the number of possibilities in half, then from 20 questions you can narrow it down from a list of a million possible characters.
Take a million and cut it in half 20 times, you end up with about 1 left.
Imagine the game like a neighborhood road system with one entrance and the road branches into two at every question, and the answer to the guess is a single house that you can reach after picking the left or right branch (yes or no answer) at each branching. On this kind of road system, if it’s 1 branching deep, you can place 2 houses; if it’s 2 branchings deep, you can place 4 houses; if it’s 3 branchings deep, you can place 8 houses; if it’s 20 branchings deep, you can place 2^(20) = 1,048,576 houses. The game just has around a million possible answers and each answer has a unique path that leads to it.
Latest Answers