def test_initialise_agent(self):
""" Test the Initialisation of the Agent Config
"""
c = Config(**dict(MODEL_NAME="TESTING"))
assert c.__repr__() == '< Player Config >'
assert isinstance(c, Config)
assert c.MODEL_NAME == "TESTING"
def create_keras_models(config1=Config(), config2=Config()):
""" Create two Keras Models for either player specifically for the TicTacToe Game
:param config1: Configuration settings for player one
:param config2: Configuration settings for player two
:return: Dictionary of Two Keras Models
"""
return {"1": KerasModel(config=config1), "-1": KerasModel(config=config2)}
def __init__(self,
name,
value,
display,
fn=None,
use_nn=False,
config=Config()):
self.name = name
self.value = value
self.display = display
self.nn = use_nn
# Get all the Config
self.c = config
if fn is not None:
self.eval_function = fn
def __init__(self, config: Config = Config()):
""" Initialise a Keras Model with its own configuration
Default will not include any ResNet blocks.
:param config:
"""
super().__init__()
if config.MODEL_TYPE == PolicyEnum.RESNET.value:
self.net = ResidualNet(config=config)
self.optimisation, self.model = self.net.compile_model()
elif config.MODEL_TYPE == "LoadingExisting":
logging.warning("Awaiting load...")
self.net = "Awaiting load..."
else:
logging.warning("Neural Network selection not known.")
raise NotImplementedError
self.config = config
def main():
""" Play two Monte Carlo Tree searches against one another...
"""
# ##########################################################
# params (Testing w/o NN):
new_game = Game(players=DEFAULT_PLAYERS)
# Play until end
while new_game.legal_plays() and new_game.winner is None:
print(new_game.player)
# Rollout the Tree
tree = MonteCarloTreeSearch(game=new_game,
evaluation_func=new_game.player.func,
node_param=new_game.player.mcts_params,
use_nn=new_game.player.use_nn)
# Run the Tree Search
tree.search(*new_game.player.mcts_search)
# Find the recommended move
# Note: Replace to use the Stochastic action selection
move, _ = tree.recommended_play(train=False)
# Stochastic: move, _ = tree.recommended_play(train=True)
# Play the recommended move and store the move
tree.show_tree(level=1)
new_game.play(move=move)
# Show the tree and board, debugging
new_game.show_board()
# ##########################################################
# params (Testing w NN):
game = Game(players=NNetPlayers, using_nn=True, nn_player=0)
models = create_keras_models(config1=Config(**{}), config2=Config(**{}))
models['1'].load_checkpoint("RLBook/Chapter8/20180214_KerasModel_TTT_V1")
# Play until end
while game.legal_plays() and game.winner is None:
print(game.player)
# Rollout the Tree
tree = MonteCarloTreeSearch(game=game,
evaluation_func=models[str(
game.player.value)].predict,
node_param=game.player.mcts_params,
use_nn=game.player.use_nn)
# Run the Tree Search
tree.search(*game.player.mcts_search)
# Find the recommended move
# Note: Replace to use the Stochastic action selection
move, _ = tree.recommended_play(train=False)
# Stochastic: move, _ = tree.recommended_play(train=True)
# Play the recommended move and store the move
tree.show_tree(level=1)
game.play(move=move)
# Show the tree and board, debugging
game.show_board()