paddle2.2.0：DQN算法训练cartpole游戏

       DQN，基于Q-learning，结合了神经网络，不再使用Q表格来存储Q值，而是用神经网络拟合的方式，可以大大减少内存的占用，同时也更加省时。

        DQN在Q-learning的基础上，又添加了经验池回放和固定Q网络两个新的技巧

LEARN_FREQ = 5
MEMORY_SIZE = 20000
MEMORY_WARMUP_SIZE = 200
BATCH_SIZE = 32
LEARNING_RATE = 0.001
GAMMA = 0.99

class Model(parl.Model):
    def __init__(self, obs_dim, act_dim):
        super().__init__()
        self.fc1 = nn.Linear(obs_dim, 128)
        self.fc2 = nn.Linear(128, 128)
        self.fc3 = nn.Linear(128, act_dim)

    def forward(self, x):
        h1 = F.relu(self.fc1(x))
        h2 = F.relu(self.fc2(h1))
        Q = self.fc3(h2)
        return Q

class Agent(parl.Agent):
    def __init__(self, algthrim, act_dim, e_greed=None, e_greed_decrement=None):
        super().__init__(algorithm)
        self.e_greed = e_greed
        self.e_greed_decrement = e_greed_decrement
        self.act_dim = act_dim
        self.global_steps = 0
        self.update_target_steps = 200

    def sample(self, obs):
        sample1 = np.random.random()
        if sample1 > self.e_greed:
            act = self.predict(obs)
        else:
            act = np.random.randint(self.act_dim)
        self.e_greed = max(0.01, self.e_greed - self.e_greed_decrement)
        return act

    def predict(self, obs):
        obs = paddle.to_tensor(obs, dtype='float32')
        predQ = self.alg.predict(obs)
        act = paddle.argmax(predQ).numpy()[0]
        return act

    def learn(self, obs, act, reward, next_obs, terminal):
        if self.global_steps % self.update_target_steps == 0:
            self.alg.sync_target()
        self.global_steps += 1

        act = np.expand_dims(act, axis=-1)
        reward = np.expand_dims(act, axis=-1)
        terminal = np.expand_dims(terminal, axis=-1)

        obs = paddle.to_tensor(obs, dtype='float32')
        act = paddle.to_tensor(act, dtype='int32')
        reward = paddle.to_tensor(reward, dtype='float32')
        next_obs = paddle.to_tensor(next_obs, dtype='float32')
        terminal = paddle.to_tensor(terminal, dtype&