The Problem of Machines Learning Go

Matthew H. Pinner

January 16, 2003

Although it has simple rules, mastering the game of Go turns out to be very difficult for a computer. The large number of possible moves, 361 at the games starts, makes for an impossibly large amount of possibilities for a brute-force search. The security of stones is difficult to evaluate and will often effect the outcome of game. Expert systems that rely of huge databases of patterns and problems become cumbersome and resistant to change. New rules often have unforeseen effects on the current knowledge. Machine learning could one day conquer the masters without a lifetime of a posteriori knowledge.

The task of providing the best move based on a given game state has yet to be fully realized. After Markus Enzenberger’s NeuroGo, the best machine learning effort thus far, trained against itself 4500 times it stated playing at level 8 in "The Many Faces of Go”. "The Many Faces of Go” is the current champion of the 21st Century Cup, a Go program competition. NeuroGo placed 6th of 14 participating programs. These computer only competitions are the closest computers come to winning any real world titles because all are still dwarfed by skilled human opponents.

I find work in this area particularly intriguing because it is becoming increasingly rare to for humans to consistently outperform machines. Enzenberger ‘s work with multiple transformations of the game state as input to the neural network seems to be applicable elsewhere. I would like to see this problem distributed amongst a large network of computers. Machines learning to master Go will be the ultimate realization of artificial intelligence.

Resources Cited: (in order of use)

    Enzenberger, Markus. The Integration of A Priori Knowledge into a Go Playing Neural Network.
    21st Century Championship Cup 2002 Results.
    The Computer Go Ladder.
    McQuade, Bryan. Machine learning and the game of Go. Master's thesis, Middlebury College, 2001.