Transcript Computer Systems Lab..

Development of a Machine-Learning-Based
AI For Go
By Justin Park
In 1997, IBM’s Deep Blue defeated
Grand Master Gary Kasparov.
The Future of AI Problem Solving:
• Artificial intelligence has been centered around
Go
• Go is an ancient Board game developed in China
from 2500-4000 years ago
• 19x19 Board Size
• Players alternate with black and white stones
• Game ends with two consecutive passes
The Challenges of Go
• Large game set
– 200-300 possible moves
– 10,000,000,000 leaves in game tree
• Difficulty in creating a heuristic function
• Pattern analysis/abstract thinking
The Solution:
(My Project)
• A machine-learning-based AI with a genetic algorithm for
“learning” new moves
• A minimalist heuristic “guiding function” for learning
basic moves
• Database storing of previously played games
• Recreation of “Roving Eye” techniques to further
adaptation to larger size boards.
Development (Python)
– Board rules: illegal moves and killing stones
– Creation of heuristic function based on
influence with respect to distance
– Sort possible moves and corresponding score
(as determined by evaluation function)
Development (continued)
• Creation of classes Game and Games to
store boards.
• Search for best move algorithm
– Comparison of boards with similar # of moves
– Heuristic function = similarity +
influence(board)
Results
Machine
Machine-Learning learns how to either win or lose
Machine-Learning function degenerates when faced
against its parent function
Machine-Learning function improves with outside
human intervention
Future Work
• Research with a larger pool of heuristic
functions
• Increase depth of heuristic search
• Compare boards with 3x3 squares
• Compatibility with GMP
• .sgf reading and writing