Tuesday, August 14, 2007

Something Incredible About The A.I. Of Chess Computers

Albert Frank Headshot by Albert Frank

I gave the following position to the best chess software, without their tablebase (which incorporates all positions with maximum six pieces on the board into their consideration).

White (to move) : Kg7, Rb1; h7
Black : Ke6, Ra8; e4

A good chess player (minimum « first category ») will immediately see that 1.Rb5 wins (Black will soon have no valid move) and that the promotion 1. h8Q gives only a draw.

Computers, after two hours (on a 3 Gz Pentium 4 with 1 G° RAM) don’t see the winning move at all, and just stay with 1. h8Q. Thus, they see (or think?) in an incredible poor way.

Once more, these computers are sometimes like “great players, stronger than any human”, and sometimes “absolutely stupid”.

I don’t know if any conclusion can be taken out of this for the moment.

1 comment:

David Michael Fabian said...

A chess program is "smart" when either (1) it is using its database (i.e. in the endgame (where it can play perfect chess if there are 5 pieces left on the board [last I knew in 1998] or in the opening), or (2) a position is open (i.e. not locked up by immobile pawns) and has many pieces that are close to fighting; but it is "stupid" otherwise. This is because it relies mostly upon brute-force searching of possibilities, which limits it to looking ahead about 10 half-moves (meaning 5 moves by each player). A human can effectively (although not necessarily precisely) look ahead many more moves in either (1) a closed position, (2) a position where he is amassing an army (which is not yet close to engaging the enemy), or (3) an endgame. This is because he can greatly reduce the number of possibilities that he must examine in order to look so far ahead in such positions. In your example, a human would reason, “If I cut off his King from his pawn, he cannot escort it to queendom,” whereas the computer would look at Rb5 and think that every possibility within the next 10 half-moves leads to just a R-vs-P endgame (assuming it did not have an endgame database of 4-piece positions)… so it plays h8Q because it thinks it has more winning possibilities (like where his opponent does not capture his new queen).