How does Akinator find specific people with such vague questions? Does it also count on luck?

254 views

How does Akinator find specific people with such vague questions? Does it also count on luck?

In: Technology

Akinator just uses a binary phylogeny tree similar to how you might identify plants of animals. So it asks a question “is your character male or female” then depending on the answer it can narrow the pool down by eliminating the impossibilities. As the pool gets smaller and smaller it can ask more specific questions, or it can ask a vague question that only has a few people in it remaining. For example the question “does your character wear glasses” is extremely broad. But since he has already asked “Is you character male, fictional, younger than 20, from a fantasy book, a wizard, parents killed by the Dark Lord” then it may have it narrowed down to only a few and the question “Does your character wear glasses?” Gives an answer.

Now when you successfully stump Akinator, it asks you to create a new question that it can use to differentiate between the answer and guess he got wrong. And by doing that it can find key words in that difference to categorize the character into another group to help narrow it down.

Because many vague questions describe one very specific person.

There’s a lot men in the world, and there’s a lot of blonde people and there’s a lot of people under 165 cm in height. But there’s really not a lot of people who are **all three** of those things, so despite all 3 questions being very vague, they end up describing a much more specific subset than any individual question on it’s own would.

Akinator just keeps asking vague questions until the subset of people who fit **all** of the answers is narrow enough to make a guess.

Have you ever played the game [*Guess Who?*](https://www.youtube.com/watch?v=a76UPzU2VXM&ab_channel=TripleSGames) It’s a game where you have to identify your opponent’s character by answering yes/no questions. You can ask something like “do they have blue eyes?” “do they have facial hair?” With each passing question, you narrow down the pool of candidates. Eventually, you eliminate everyone but one.

Akinator works the same way, except it’s a much large pool of candidates. However, they use user input to help them. Let’s say that you pick a character that nobody else has. Akinator will remember all your answers, and should someone think of that character at another time, Akinator will use your answers to help guide them to the correct solution. The more often people use the same character, the more refined the path the solution can get.

If you ask the right questions – questions that continually split the list of all possible people in half – you can uniquely identify any human on Earth in just 33 questions.

The questions therefore should be vague (that is, applies to about half of everyone) so that you eliminate as many possibilities in one question as possible, no matter what the answer is. If it’s too specific, the chance of the answer being “yes” is much smaller, and to worsen the matter, if the answer is “no”, then you only eliminated a small number of people with your question.

100 people.
1 question, down to 50 people.
2nd question down to 25.
3rd, down down to 10
4th, down to 6.
At this point it likely realized the cost of returns for any question is lacking. It likely takes a guess based on the percentage of visitors that end up there, or if the odds are low enough asks another question to eliminate at least 1 possibility.

To go along with all the algorithms and math it does one other thing. Every time someone thinks of a person/character/whatever it records the questions and answers that got it there and adds them to the path for that specific result. Doesn’t always work but when 50,000 people all think of Goku it gets easier.

Goku: From an anime, black hair, male, transforms, fights, etc

I get the math behind it but some times the thing will guess with pinpoint accueacry as if it knew.

Ah yes, male cartoon character

In an anime?

Do they fight with cards?

Oh could they be from Yu gi Oh?

Bam , Yu Gi Moto.

Next time , me thinking of Sora

Ah, male character.

Do they appear in a cartoon/video game series?

Before even asking for hair color or any trait. **Would you say this character fights with a Keyblade^TM**?

Akinator is struggling with Ace Ventura right now. 80 questions in and it guessed Earnest P Worl, and Bob from what about Bob.

But what is it using to find all of these people? What database and info is it pulling from for tens of thousands of people

It used probability and statistics and binary search using weighted inputs, along with rudimentary AI for knowing what questions to ask based on prior players playing.

Since everyone is familiar with the concept of a family tree, Imagine a similar tree structure, but one which consists of random people connected together.

They have been placed according to certain patterns like males on the left of the tree, females on the right at high level and then scientists on the left and teachers on the right at next level and so on.

The questions that you see on the screen is typically the logic applied at that node in a tree. Every question that you answer is used to select either a left or right part of the tree and you descend down the tree with your answer. Once the end of the tree is reached, we get answer we are looking for.

The only famous person I was ever able to trip it up on was John Waters of Hairspray! fame. I answered every question right and it still couldn’t get it. This was a short while back and I filled in who I was thinking of in the “You’ve defeated me” page, so it’s probably better now. (It gave me James Cameron).