120 likes | 234 Views
Go. Rules of Go. Played on a 19 x 19 board 2 players alternate turns On a turn a player can place a piece or pass Once all a piece’s liberties have been captured, the piece is removed The game ends when both players pass consecutively. Capturing a piece.
E N D
Rules of Go • Played on a 19 x 19 board • 2 players alternate turns • On a turn a player can place a piece or pass • Once all a piece’s liberties have been captured, the piece is removed • The game ends when both players pass consecutively
Capturing a piece • The black piece shown has one remaining liberty • If white were to go in spot A, black would have no liberties left and the black piece would be captured
Other rules • Komi • A handicap given to the player who goes second, generally 5½ points • Ko • A player cannot recreate a previous game layout
Difficulties • Game tree is astronomically big • Estimated to be about 10170 • Branching factor • On average about 235 (versus 35 for chess) • Game length • Can be up to 300 moves • Passing
More Difficulties • Evaluation function • Very hard to come up with a good one • No or poor Alpha-Beta pruning • All pieces have the same value • The number of pieces on the board have no correlation to the potential outcome of the game • This means in order to have a simple evaluation heuristic, the program must use brute force to traverse down the tree
Even More Difficulties • Humans can look ahead very far in Go • Some situations are very pattern-oriented • Difficult to tell when program doesn’t need to look far down the search tree • Grouping of pieces • Until the end of the game, relating pieces together is very difficult • If veteran Go players find it difficult, imagine how hard this reasoning is for a program
Go Program Techniques • Simplifying the game • Splitting the board into sections • Focusing on the strategies of specific groups of pieces • By reducing the game into sub-problems, the program can use Go techniques coupled with searching smaller trees
Go Program Techniques • Memorizing patterns and optimal moves • Certain moves in certain situations guarantee the capture of pieces • When these patterns occur in the game, no tree searching is required • Joseki database
The Many Faces of Go • Uses local searches on parts of the board to determine if a group of pieces cannot be captured or if nothing can prevent them form being captured • Rated on par with the average human player who has one year of experience • Won the Ing cup in 1998
The Many Faces of Go • Uses many different types of tree searches • Each type of search tries to accomplish something different • Connecting two groups • Capturing a group • Making a group dead or alive
Improvement Still Needed • Program playing style can be easily beaten • The Many Faces of Go lost to an expert even with a 29 stone handicap • Cash Prize