In this project, we will build a modified hybrid of AlphaGo and AlphaZero for 9x9 Go from almost scratch (using only computational libraries).
| Week | Links |
|---|---|
| 1 | Slides, Play Online Go, Computer Go Rules |
| 2 | Slides, NN Arch, PyTorch CNN example |
| 3 | Slides, PyTorch Modules |
| 4 | Slides, SL Description |
| 5 | Slides |
For a more detailed list of topics and resources, check the most recent "This Week in Mini-AlphaGo" email (released every Wednesday afternoon).
board.py contains the Board class, which is a custom Go board interface written specifically for this project. Board handles all the game logic of Go, including moves, captures, end detection, and scoring. The file contains example usage and docstrings for each function.
It is important to note that Board objects contain a .grid, which is a np.NDArray with shape (size, size) containing ints that are 0 (empty), 1 (black stone), or 2 (white stone).
GameNode inherits from Board and implements a tree structure with .prev (the GameNode leading to self), .prev_move (the move tuple that goes from .prev to self), and .nexts (a possibly incomplete list of child GameNodes).
Examples
Count the number of black stones on a Board after playing some moves
board = Board(9) # Init empty 9x9 Board
board.play_stone(5, 6) # Black plays at (5, 6)
board.play_stone(0, 1) # White plays at (0, 1)
board.play_stone(0, 0) # Black plays at (0, 0)
print(board) # Visual board representation
print(f"Black stones: {(board.grid == 1).sum()}") # 2Print history leading up to a GameNode
node = GameNode(9) # Init empty 9x9 GameNode
node = node.create_child((5, 6)) # Black plays at (5, 6)
node = node.create_child((0, 1)) # White plays at (0, 1)
node = node.create_child((0, 0)) # Black plays at (0, 0)
# Print game history in reverse
while node != None:
print(node)
node = node.prevelo_calculator.py, contains the Elo_calculator class, which allows you to play to Bots against each other, calculate their updated elos, and store them.
bot.py has the basic structure for the Bot class, which will be an abstraction for the Go bots we will build that provides a consistent interface for evaluation.
In the following code:
go = Board(9)
elo = Elo_calculator(game=go, prev_elo_data = "elo_data.pkl")
random_player1 = Bot()
random_player2 = Bot()
elo.register_bot(name= "bob", bot = random_player1)
elo.register_bot(name = "bobette", bot = random_player2)
elo.play_match("bob", "bobette")
elo.save("elo_data.pkl")We create an elo_calculator object from past elo data, register 2 agents, and run them against each other, saving the data to 'elo_data.pkl'. If you were to run this code again, then the code below would do the same this as the code snippet above:
go = Board(9)
elo = Elo_calculator(game=go, prev_elo_data = "elo_data.pkl")
elo.play_match("bob", "bobette")
elo.save("elo_data.pkl")