I've just finished watching this video. Human Computation, a lecture about using simple web based games to produce high quality image clasification data through asymetric and symetric varification. I was wondering if anyone is aware of any games in the past that have harvested data from individual users in order to improve gameplay for the wider audience. Do you think it would be correct to say that as long as for a given problem you have a way of recording/quantifying that users input or behavior, that you can then incorporate this into a game in such a way that other players would in effect varify its correctness? And through this varification the computer can learn well? Do any games use the internet to share data to improve the games artifical intellence? Or is the learning process still too unpredictable to be useful?

Quote: Original post by aleks_1661
I was wondering if anyone is aware of any games in the past that have harvested data from individual users in order to improve gameplay for the wider audience.

Just about every major Chess program? All their big move-dictionaries are based on past games of grandmasters and whatnot.

Quote: Original post by aleks_1661
Do you think it would be correct to say that as long as for a given problem you have a way of recording/quantifying that users input or behavior, that you can then incorporate this into a game in such a way that other players would in effect varify its correctness? And through this varification the computer can learn well?

One of my AI proff's once mentioned the idea of machine learning via Imitation of human behaviours (rather than trying to do minimax searches and whatnot). I've also seen this idea used in the first chapter of the book 'Berserker' (by fred saberhagen)
So yeah, why not.

Quote: Original post by aleks_1661
Do you think it would be correct to say that as long as for a given problem you have a way of recording/quantifying that users input or behavior, that you can then incorporate this into a game in such a way that other players would in effect varify its correctness? And through this varification the computer can learn well?

If you can incorporate specific behaviours into a game playing agent (either by watching someone else, being taught or trial and error) then playing the game against that player is a way to validate which of two behaviours is more likely to result in a win. This gives you a relative ranking of the competing behaviours. If you play that same behaviour over a broad spectrum of other behaviours, then you find where it fits within the population of possible behaviours. This applies equally to human players and machine players. One good thing about game theory is that you don't need absolute utilities but rather just relative utilities to enable differentiation of strategy.

Cheers,

Timkin