ddpg.py 文件源码

def test_ddpg():
    import gym_mix
    env = gym.make('ContinuousCopyRand-v0')
    env = wrappers.TimeLimit(env, max_episode_steps=0)

    @model(optimizer=tf.train.AdamOptimizer(0.0001),
                 tracker=tf.train.ExponentialMovingAverage(1 - 0.001))
    def actor(x):
        x = layers.fully_connected(x, 50, biases_initializer=layers.xavier_initializer())
        a = layers.fully_connected(x, env.action_space.shape[0], None,
                                                             weights_initializer=tf.random_normal_initializer(0, 1e-4))
        return a

    @model(optimizer=tf.train.AdamOptimizer(.001),
                 tracker=tf.train.ExponentialMovingAverage(1 - 0.001))
    def critic(x, a):
        x = layers.fully_connected(x, 300, biases_initializer=layers.xavier_initializer())
        x = tf.concat([x, a], axis=1)
        x = layers.fully_connected(x, 300, biases_initializer=layers.xavier_initializer())
        x = layers.fully_connected(x, 300, biases_initializer=layers.xavier_initializer())
        q = layers.fully_connected(x, 1, None, weights_initializer=tf.random_normal_initializer(0, 1e-4))
        return tf.squeeze(q, 1)

    agent = DdpgAgent(env, actor, critic)

    for ep in range(10000):
        R, _ = agent.play_episode()

        if ep % 100 == 0:
            print(f'Return after episode {ep} is {R}')