reinforce.py 文件源码

def __init__(self, env, monitor_path, video=True, **usercfg):
        super(REINFORCE, self).__init__(**usercfg)
        self.env = wrappers.Monitor(env, monitor_path, force=True, video_callable=(None if video else False))
        self.env_runner = EnvRunner(self.env, self, usercfg)
        self.monitor_path = monitor_path
        # Default configuration. Can be overwritten using keyword arguments.
        self.config.update(dict(
            batch_update="timesteps",
            timesteps_per_batch=1000,
            n_iter=100,
            gamma=0.99,  # Discount past rewards by a percentage
            decay=0.9,  # Decay of RMSProp optimizer
            epsilon=1e-9,  # Epsilon of RMSProp optimizer
            learning_rate=0.05,
            n_hidden_units=20,
            repeat_n_actions=1,
            save_model=False
        ))
        self.config.update(usercfg)

        self.build_network()
        self.make_trainer()

        init = tf.global_variables_initializer()
        # Launch the graph.
        self.session = tf.Session()
        self.session.run(init)
        if self.config["save_model"]:
            tf.add_to_collection("action", self.action)
            tf.add_to_collection("states", self.states)
            self.saver = tf.train.Saver()
        self.rewards = tf.placeholder("float", name="Rewards")
        self.episode_lengths = tf.placeholder("float", name="Episode_lengths")
        summary_loss = tf.summary.scalar("Loss", self.summary_loss)
        summary_rewards = tf.summary.scalar("Rewards", self.rewards)
        summary_episode_lengths = tf.summary.scalar("Episode_lengths", self.episode_lengths)
        self.summary_op = tf.summary.merge([summary_loss, summary_rewards, summary_episode_lengths])
        self.writer = tf.summary.FileWriter(os.path.join(self.monitor_path, "task0"), self.session.graph)