What I want is something different: mining tons of games to learn new stuff about positional strength in chess. For example, assume that we don't know that a queen is stronger than a knight, and instead think that the two pieces have the same value. We would trade those two pieces more or less indifferently, resulting in lots of games where one side has a queen and the other side has a knight. Now, the program I would like to do will mine all those games and say "99% of the cases where one side had a queen and the other side had a knight were a win for the side with the queen". And we would be able to learn that a queen is stronger to a knight... Now obviously this example is too simple, but I would like to see if it is possible to learn some more subtle things.
Re:What are some fun project ideas in machine learning? :)
by keefe on Saturday February 09, 2013 @02:22PM
(#5193632)
Have you checked out chess problems and go problems? They're usually presented in bulk in structured data.
http://www.chessproblems.com/
I haven't look at any in a long time, but it at least should highlight some problems hard for humans, which implies something can be learned there, imho.
[ Parent
| on HN ]
Re:What are some fun project ideas in machine learning? :)
by S4M on Thursday February 14, 2013 @04:27AM
(#5218640)
Thanks for the link, but that's not exactly what I wanted. The chess problem you are showing are some composition problem: someone creating a position where it is possible to checkmate in two moves, but the moves are very unintuitive and the position is completely unnatural, which is very hard for a human player to find the solution. A computer, on the other hand, will have no problem to find it by brute force.