ddqn.py 文件源码

def __init__(self):
        print "Initializing DQN..."
        self.exploration_rate = config.rl_initial_exploration
        self.fcl_eliminated = True if len(config.q_fc_hidden_units) == 0 else False

        # Q Network
        conv, fc = build_q_network(config)
        self.conv = conv
        if self.fcl_eliminated is False:
            self.fc = fc
        self.load()
        self.update_target()

        # Optimizer
        ## RMSProp, ADAM, AdaGrad, AdaDelta, ...
        ## See http://docs.chainer.org/en/stable/reference/optimizers.html
        self.optimizer_conv = optimizers.Adam(alpha=config.rl_learning_rate, beta1=config.rl_gradient_momentum)
        self.optimizer_conv.setup(self.conv)
        if self.fcl_eliminated is False:
            self.optimizer_fc = optimizers.Adam(alpha=config.rl_learning_rate, beta1=config.rl_gradient_momentum)
            self.optimizer_fc.setup(self.fc)

        # Replay Memory
        ## (state, action, reward, next_state, episode_ends)
        shape_state = (config.rl_replay_memory_size, config.rl_agent_history_length * config.ale_screen_channels, config.ale_scaled_screen_size[1], config.ale_scaled_screen_size[0])
        shape_action = (config.rl_replay_memory_size,)
        self.replay_memory = [
            np.zeros(shape_state, dtype=np.float32),
            np.zeros(shape_action, dtype=np.uint8),
            np.zeros(shape_action, dtype=np.int8),
            np.zeros(shape_state, dtype=np.float32),
            np.zeros(shape_action, dtype=np.bool)
        ]
        self.total_replay_memory = 0
        self.no_op_count = 0