return VanillaPolicyGradient("model", discount_factor,
learning_rate_policy, learning_rate_advantage,
value_steps_per_update, config,
lambda_parameter)
def _define_agent(model: VanillaPolicyGradient) -> VPGRNGDiscreteAgent:
# Define attributes
updates_per_training_volley: int = 2
# Return the agent
return VPGRNGDiscreteAgent("vpg_agent", model, updates_per_training_volley)
if __name__ == "__main__":
# Parse the command line arguments
workspace_path, experiment_iterations_number, cuda_devices, render_during_training, render_during_validation, render_during_test = command_line_parse(
)
# Define the CUDA devices in which to run the experiment
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = cuda_devices
# Define the logger
logger: logging.Logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
# Define Neural Network layers
nn_config: Config = Config()
nn_config.add_hidden_layer(LayerType.dense, [
4096, tensorflow.nn.relu, True,
tensorflow.contrib.layers.xavier_initializer()
])
nn_config.add_hidden_layer(LayerType.dense, [
4096, tensorflow.nn.relu, True,
tensorflow.contrib.layers.xavier_initializer()
exploration_rate_decay: float = 0.00002
# Return the explorer
return EpsilonGreedyExplorationPolicy(exploration_rate_max, exploration_rate_min, exploration_rate_decay)
def _define_epsilon_greedy_agent(model: DuelingDeepQLearning, exploration_policy: EpsilonGreedyExplorationPolicy) -> DDDQLTicTacToeAgent:
# Define attributes
weight_copy_step_interval: int = 100
batch_size: int = 150
# Return the agent
return DDDQLTicTacToeAgent("dddqn_egreedy_agent", model, exploration_policy, weight_copy_step_interval, batch_size)
if __name__ == "__main__":
# Parse the command line arguments
checkpoint_path, iteration_number, cuda_devices, render = command_line_parse(True)
# Define the CUDA devices in which to run the experiment
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = cuda_devices
# Define the logger
logger: logging.Logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
# Tic Tac Toe random environment:
# - success threshold to consider both the training completed and the experiment successful is around 95% of match won by the agent (depending on reward assigned)
environment_name: str = 'TicTacToeRandom'
# Generate Tic Tac Toe environments with random environment player and using the O player as the environment player only with low reward type
environment_low_reward: TicTacToeEnvironmentRandom = TicTacToeEnvironmentRandom(environment_name, Player.o,
1.0, -0.1, 0.0)
# Define Neural Network layers
nn_config: Config = Config()
nn_config.add_hidden_layer(LayerType.dense, [1024, tensorflow.nn.relu, True, tensorflow.contrib.layers.xavier_initializer()])
