dqn_trainer.py 文件源码

def __init__(self, 
                    agent, 
                    memory_size=10**4,
                    replay_size=32,
                    gamma=0.99,
                    initial_exploration=10**4,
                    target_update_freq=10**4,
                    learning_rate=0.00025,
                    epsilon_decay=1e-6,
                    minimum_epsilon=0.1):
        self.agent = agent
        self.target = Q(self.agent.q.n_history, self.agent.q.n_action, on_gpu=self.agent.q.on_gpu)

        self.memory_size = memory_size
        self.replay_size = replay_size
        self.gamma = gamma
        self.initial_exploration = initial_exploration
        self.target_update_freq = target_update_freq
        self.learning_rate = learning_rate
        self.epsilon_decay = epsilon_decay
        self.minimum_epsilon = minimum_epsilon
        self._step = 0

        # prepare memory for replay
        n_hist = self.agent.q.n_history
        size = self.agent.q.SIZE
        self.memory = [
            np.zeros((memory_size, n_hist, size, size), dtype=np.float32),
            np.zeros(memory_size, dtype=np.uint8),
            np.zeros((memory_size, 1), dtype=np.float32),
            np.zeros((memory_size, n_hist, size, size), dtype=np.float32),
            np.zeros((memory_size, 1), dtype=np.bool)
        ]
        self.memory_text = [
            "state", "action", "reward", "next_state", "episode_end"
        ]

        # prepare optimizer
        self.optimizer = optimizers.RMSpropGraves(lr=learning_rate, alpha=0.95, momentum=0.95, eps=0.01)
        self.optimizer.setup(self.agent.q)
        self._loss = 9
        self._qv = 0