Sorry this isn`t really a question about AI but it seemed the most appropriate category to put it in. I had an idea for testing collective intelligence, it seemed so obvious I wondered if perhaps anyone had heard of it already being tried? It would be to setup a multiplayer game of (for example) chess. The human players would be divided into teams and unable to confer, and would vote on the move to take the most popular move being the one taken. Has anyone heard of this being attempted before?

Yes, try searching for Kasparov vs The World.

I've got a book on my bookshelf called "Out Of Control" by Kevin Kelly. It's essentially a book about collective intelligence. There's been heaps of research in this area and one of my favourites is the work of Loren Carpenter and the games he plays with human audiences! ;)

I would think that a Turing test for human intelligence would be interesting. There's Person A, Person B, and Group C. B tries to figure out who is the group and who is the person. The group votes or something to decide what answers they give. C tries to make B guess incorrectly, while A tries to get B to guess correctly.

Could B guess it correctly? Has anyone already done this?

Thats how google works. Whats the best page on any subjet? Lets ask millions of other web pages and take the most popular answer.

Thats why its better at spell-checking than anything else, too.

Some fuzzy logic systems also works like that. By asking the same question to different systems and combining the answers.

Oh and there are several game shows on TV where players can ask the audience. But honestly, the audience is often dumb. Isnt that what Family Feud was about? Being an expert on the question asked was an hindrance in that quiz show, because the general public isnt that smart on average.

Another book of interest on this subject is the The Wisdom of Crowds.

It is written from a sociological perspective, but the conclusions are similar (nearly identical) to those in the current field of ensemble methods in pattern classification and statistical learning.

-Kirk

The stock market is also specifically designed around collective intelligence.

As are the new captcha systems where you have to solve 2 capthas: one is a known one, the other is an unknown piece of text scanned from a book that OCR systems cannot decipher. Once enough people agree on what the unknown captcha means, it is removed from the unknown list and the scan of the book is updated.

Various intelligence agencies have also played around with making stock markets for intelligence gathering. i.e. you buy shares of "a terrorist attack will happen on X date". mostly those have been shut down because of their questionable morality =)

I don't know of a "game" game to which this has been applied.

-me

Genetic AI algorithms are also sort of collective intelligence. They combine all the ideas, input, traits, whatever from whatever sources (can be multiple external sources) then either rate them percentage wise according to an algorithm, or simply give them dominant or recessive trait, then combine them all to find the best answer.

The theory then is that the two best answers "mate" as in reproduce and have a "child" that is meant to be a hybrid that combines all the best traits of both parents (dominant traits).

It's meant to simulate real world evolving but it's actually fatally flawed because in the real world the two "best" of anything do not always or even usually combine. AI genetic algorithm engineers try to get around this by introducing a "mutation" algorithm which is just a fancy name for inserting a random seed periodically to offset the fatal flaw in the original algorithm.

How do you define "best" anyway? It's completely subjective and open to limitless interpretation. Which is the real basis for genetics anyway... infinite possibilities of combination, not a pre-determined selection. Seems to me that it's all back to front.

m0ng00se

Quote:Original post by m0ng00se
Genetic AI algorithms are also sort of collective intelligence. They combine all the ideas, input, traits, whatever from whatever sources (can be multiple external sources) then either rate them percentage wise according to an algorithm, or simply give them dominant or recessive trait, then combine them all to find the best answer.

The theory then is that the two best answers "mate" as in reproduce and have a "child" that is meant to be a hybrid that combines all the best traits of both parents (dominant traits).

The usual theory is that the worst few answers are killed off (half or so), the best answers are combined arbitrarily (non-gender related - you don't get half "female" answers and half "male" answers etc.) and generate a new population. You don't take the best two since that's a very flawed approach.

Quote:
It's meant to simulate real world evolving but it's actually fatally flawed because in the real world the two "best" of anything do not always or even usually combine. AI genetic algorithm engineers try to get around this by introducing a "mutation" algorithm which is just a fancy name for inserting a random seed periodically to offset the fatal flaw in the original algorithm.

The mutation is to make sure local highs don't prevent reaching the global high. The idea behind that is that there are concepts that are so good that going in any direction will produce a worse outcome - yet, if you go in a particular direction long enough it shows to be a higher point than the original one. This creates species.

Also, as populations evolve their symbol spread decreases making certain possibilities impossible and stagnating the population to within the bounds of their current optimum. Mutation allows arbitrary changes allowing them to go beyond those bounds. If you would disallow mutation in groups, any change after a while of stagnance would make the group disappear as it doesn't have the genetic language to define a new answer.

Quote:
How do you define "best" anyway? It's completely subjective and open to limitless interpretation. Which is the real basis for genetics anyway... infinite possibilities of combination, not a pre-determined selection. Seems to me that it's all back to front.

Locally, of course. If you want to live on the poles, you'd be judged by your ability to withstand cold and to find food, plus how to reproduce and raise a kid there. If you're an answer to a given question, you will be rated to multiple effects of that answer to see how good the answer is.

You don't apply this to yes/no questions because the only evaluation would be either perfect or useless.