Prioritized Experience Replay (DQN) (Tensorflow)

class SumTree(object):
    # 建立 tree 和 data,
    # 因为 SumTree 有特殊的数据结构,
    # 所以两者都能用一个一维 np.array 来存储
    def __init__(self, capacity):

    # 当有新 sample 时, 添加进 tree 和 data
    def add(self, p, data):

    # 当 sample 被 train, 有了新的 TD-error, 就在 tree 中更新
    def update(self, tree_idx, p):

    # 根据选取的 v 点抽取样本
    def get_leaf(self, v):

    # 获取 sum(priorities)
    @property
    def totoal_p(self):

class Memory(object):
    # 建立 SumTree 和各种参数
    def __init__(self, capacity):

    # 存储数据, 更新 SumTree
    def store(self, transition):

    # 抽取 sample
    def sample(self, n):

    # train 完被抽取的 samples 后更新在 tree 中的 sample 的 priority
    def batch_update(self, tree_idx, abs_errors):

class DQNPrioritiedReplay:
    def __init__(..., prioritized=True, sess=None)
        self.prioritized = prioritized
        ...
        if self.prioritized:
            self.memory = Memory(capacity=memory_size)
        else:
            self.memory = np.zeros((self.memory_size, n_features*2+2))

        if sess is None:
            self.sess = tf.Session()
            self.sess.run(tf.global_variables_initializer())
        else:
            self.sess = sess

class DQNPrioritizedReplay:
    def _build_net(self)
        ...
        # self.prioritized 时 eval net 的 input 多加了一个 ISWeights
        self.s = tf.placeholder(tf.float32, [None, self.n_features], name='s')  # input
        self.q_target = tf.placeholder(tf.float32, [None, self.n_actions], name='Q_target')  # for calculating loss
        if self.prioritized:
            self.ISWeights = tf.placeholder(tf.float32, [None, 1], name='IS_weights')

        ...
        # 为了得到 abs 的 TD error 并用于修改这些 sample 的 priority, 我们修改如下
        with tf.variable_scope('loss'):
            if self.prioritized:
                self.abs_errors = tf.reduce_sum(tf.abs(self.q_target - self.q_eval), axis=1)    # for updating Sumtree
                self.loss = tf.reduce_mean(self.ISWeights * tf.squared_difference(self.q_target, self.q_eval))
            else:
                self.loss = tf.reduce_mean(tf.squared_difference(self.q_target, self.q_eval))

class DQNPrioritizedReplay:
    def store_transition(self, s, a, r, s_):
        if self.prioritized:    # prioritized replay
            transition = np.hstack((s, [a, r], s_))
            self.memory.store(transition)
        else:       # random replay
            if not hasattr(self, 'memory_counter'):
                self.memory_counter = 0
            transition = np.hstack((s, [a, r], s_))
            index = self.memory_counter % self.memory_size
            self.memory[index, :] = transition
            self.memory_counter += 1

class DQNPrioritizedReplay:
    def learn(self):
        ...
        # 相对于 DQN 代码, 改变的部分
        if self.prioritized:
            tree_idx, batch_memory, ISWeights = self.memory.sample(self.batch_size)
        else:
            sample_index = np.random.choice(self.memory_size, size=self.batch_size)
            batch_memory = self.memory[sample_index, :]

        ...

        if self.prioritized:
            _, abs_errors, self.cost = self.sess.run([self._train_op, self.abs_errors, self.loss],
                                         feed_dict={self.s: batch_memory[:, :self.n_features],
                                                    self.q_target: q_target,
                                                    self.ISWeights: ISWeights})
            self.memory.batch_update(tree_idx, abs_errors)   # update priority
        else:
            _, self.cost = self.sess.run([self._train_op, self.loss],
                                         feed_dict={self.s: batch_memory[:, :self.n_features],
                                                    self.q_target: q_target})

        ...