class DefaultAlphaZeroPlayer(Player):
def __init__(self,
game,
arg_file: typing.Optional[str] = None,
name: str = "") -> None:
super().__init__(game, arg_file, name, parametric=True)
if self.player_args is not None:
self.args = DotDict.from_json(self.player_args)
self.model = DefaultAlphaZero(self.game, self.args.net_args,
self.args.architecture)
self.search_engine = AlphaZeroMCTS(self.game, self.model,
self.args.args)
self.name = self.args.name
def set_variables(self, model, search_engine, name):
self.model = model
self.search_engine = search_engine
self.name = name
def refresh(self, hard_reset: bool = False):
super().refresh()
self.search_engine.clear_tree()
def act(self, state: GameState) -> int:
pi, _ = self.search_engine.runMCTS(state, self.history, temp=0)
return np.argmax(pi).item()
class DefaultAlphaZeroPlayer(Player):
"""
Standard AlphaZero agent that samples actions from MCTS within a given environment model.
"""
def __init__(self,
game,
arg_file: typing.Optional[str] = None,
name: str = "") -> None:
super().__init__(game, arg_file, name, parametric=True)
if self.player_args is not None:
# Initialize AlphaZero by loading its parameter config and constructing the network and search classes.
self.args = DotDict.from_json(self.player_args)
self.model = DefaultAlphaZero(self.game, self.args.net_args,
self.args.architecture)
self.search_engine = AlphaZeroMCTS(self.game, self.model,
self.args.args)
self.name = self.args.name
def set_variables(self, model, search_engine, name: str) -> None:
""" Assign Neural Network and Search class to an external reference """
self.model = model
self.search_engine = search_engine
self.name = name
def refresh(self, hard_reset: bool = False):
""" Refresh internal state of the Agent along with stored statistics within the MCTS tree """
super().refresh()
self.search_engine.clear_tree()
def act(self, state: GameState) -> int:
""" Sample actions using MCTS using the given environment model. """
pi, _ = self.search_engine.runMCTS(state, self.history, temp=0)
return np.argmax(pi).item()
def __init__(self, exploration, budget, model, boardsize, temperature=0):
"""
The AlphaZero algorihtm must be configured with the parameters for the
PUCT formula for childnode selection and expansion along with a search budget.
Also the neural network must be provided along with the boardsize in order
to initialize the MCTS backend of AlphaZero.
:param exploration: float The exploration parameter of PUCT (c_puct).
:param budget: int The amount of MCTS simulations to perform (N).
:param model: AlphaZeroModel.Model A AlphaZeroModel backend AlphaZeroModel that guides MCTS's search
:param boardsize: int Size of the game board in order to initialize MCTS.
:param temperature: int Governs the degree of exploration. (0 = greedy)
"""
super().__init__()
self.args = dotdict({'numMCTSSims': budget, 'cpuct': exploration})
self.model = model
self.game = HexGame(boardsize)
self.searcher = MCTS(self.game, self.model, self.args)
self.temperature = temperature # temp=0 implies greedy actions
def __init__(self,
game,
arg_file: typing.Optional[str] = None,
name: str = "") -> None:
super().__init__(game, arg_file, name, parametric=True)
if self.player_args is not None:
self.args = DotDict.from_json(self.player_args)
self.model = DefaultAlphaZero(self.game, self.args.net_args,
self.args.architecture)
self.search_engine = AlphaZeroMCTS(self.game, self.model,
self.args.args)
self.name = self.args.name
class AlphaZeroPolicy(Policy):
"""
Derived class of Policy to choose actions based on the AlphaZero Algorithm.
"""
def __init__(self, exploration, budget, model, boardsize, temperature=0):
"""
The AlphaZero algorihtm must be configured with the parameters for the
PUCT formula for childnode selection and expansion along with a search budget.
Also the neural network must be provided along with the boardsize in order
to initialize the MCTS backend of AlphaZero.
:param exploration: float The exploration parameter of PUCT (c_puct).
:param budget: int The amount of MCTS simulations to perform (N).
:param model: AlphaZeroModel.Model A AlphaZeroModel backend AlphaZeroModel that guides MCTS's search
:param boardsize: int Size of the game board in order to initialize MCTS.
:param temperature: int Governs the degree of exploration. (0 = greedy)
"""
super().__init__()
self.args = dotdict({'numMCTSSims': budget, 'cpuct': exploration})
self.model = model
self.game = HexGame(boardsize)
self.searcher = MCTS(self.game, self.model, self.args)
self.temperature = temperature # temp=0 implies greedy actions
def generate_move(self, hex_board):
"""
With the given parameters for the AlphaZero procedure and a currently
provided game-state (=hex_board) call the backend MCTS class to
perform forward search guided by the neural network to choose
a move for the current player.
:param hex_board: HexBoard Class for game-logic.
:return: tuple Coordinate on the HexBoard to move to.
:see: MCTSSearcher from .hex_search
"""
self.calls += 1
if self.perspective is None:
raise Exception("No search/ player perspective given.")
# The neural network AlphaZeroModel only learns in an uniform player perspective.
# Hence we first alter the symmetry of the board according to the current player.
# The returned move is of course transformed to the perspective of the
# board that was provided in the function's argument.
search_board = self.game.getCanonicalForm(np.copy(hex_board.board),
self.perspective)
pi = self.searcher.runMCTS(search_board, temp=self.temperature)
move_idx = np.argmax(pi)
move = (move_idx // hex_board.size, move_idx % hex_board.size)
if self.perspective == -1: # Canonical form finds a move on a transposed board.
move = move[::-1]
return move
def __init__(self,
game,
arg_file: typing.Optional[str] = None,
name: str = "") -> None:
super().__init__(game, arg_file, name, parametric=True)
if self.player_args is not None:
# Initialize AlphaZero by loading its parameter config and constructing the network and search classes.
self.args = DotDict.from_json(self.player_args)
self.model = DefaultAlphaZero(self.game, self.args.net_args,
self.args.architecture)
self.search_engine = AlphaZeroMCTS(self.game, self.model,
self.args.args)
self.name = self.args.name