python - 具有独立 Action 空间的软 Actor 评论家

我正在尝试为离散的动作空间实现软角色评论器算法，并且我对损失函数有麻烦。

这是SAC与连续动作空间的链接：
https://spinningup.openai.com/en/latest/algorithms/sac.html

我不知道我在做什么错。

问题是网络在柱极环境上没有学到任何东西。

github上的完整代码：https://github.com/tk2232/sac_discrete/blob/master/sac_discrete.py

这是我的想法，如何计算离散操作的损失。

价值网络

class ValueNet:
    def __init__(self, sess, state_size, hidden_dim, name):
        self.sess = sess

        with tf.variable_scope(name):
            self.states = tf.placeholder(dtype=tf.float32, shape=[None, state_size], name='value_states')
            self.targets = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='value_targets')
            x = Dense(units=hidden_dim, activation='relu')(self.states)
            x = Dense(units=hidden_dim, activation='relu')(x)
            self.values = Dense(units=1, activation=None)(x)

            optimizer = tf.train.AdamOptimizer(0.001)

            loss = 0.5 * tf.reduce_mean((self.values - tf.stop_gradient(self.targets)) ** 2)
            self.train_op = optimizer.minimize(loss, var_list=_params(name))

    def get_value(self, s):
        return self.sess.run(self.values, feed_dict={self.states: s})

    def update(self, s, targets):
        self.sess.run(self.train_op, feed_dict={self.states: s, self.targets: targets})

在Q_Network中，我使用收集的操作收集值

例

q_out = [[0.5533, 0.4444], [0.2222, 0.6666]]
collected_actions = [0, 1]
gather = [[0.5533], [0.6666]]

收集功能

def gather_tensor(params, idx):
    idx = tf.stack([tf.range(tf.shape(idx)[0]), idx[:, 0]], axis=-1)
    params = tf.gather_nd(params, idx)
    return params

Q网

class SoftQNetwork:
    def __init__(self, sess, state_size, action_size, hidden_dim, name):
        self.sess = sess

        with tf.variable_scope(name):
            self.states = tf.placeholder(dtype=tf.float32, shape=[None, state_size], name='q_states')
            self.targets = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='q_targets')
            self.actions = tf.placeholder(dtype=tf.int32, shape=[None, 1], name='q_actions')

            x = Dense(units=hidden_dim, activation='relu')(self.states)
            x = Dense(units=hidden_dim, activation='relu')(x)
            x = Dense(units=action_size, activation=None)(x)
            self.q = tf.reshape(gather_tensor(x, self.actions), shape=(-1, 1))

            optimizer = tf.train.AdamOptimizer(0.001)

            loss = 0.5 * tf.reduce_mean((self.q - tf.stop_gradient(self.targets)) ** 2)
            self.train_op = optimizer.minimize(loss, var_list=_params(name))

    def update(self, s, a, target):
        self.sess.run(self.train_op, feed_dict={self.states: s, self.actions: a, self.targets: target})

    def get_q(self, s, a):
        return self.sess.run(self.q, feed_dict={self.states: s, self.actions: a})

政策网

class PolicyNet:
    def __init__(self, sess, state_size, action_size, hidden_dim):
        self.sess = sess

        with tf.variable_scope('policy_net'):
            self.states = tf.placeholder(dtype=tf.float32, shape=[None, state_size], name='policy_states')
            self.targets = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='policy_targets')
            self.actions = tf.placeholder(dtype=tf.int32, shape=[None, 1], name='policy_actions')

            x = Dense(units=hidden_dim, activation='relu')(self.states)
            x = Dense(units=hidden_dim, activation='relu')(x)
            self.logits = Dense(units=action_size, activation=None)(x)
            dist = Categorical(logits=self.logits)

            optimizer = tf.train.AdamOptimizer(0.001)

            # Get action
            self.new_action = dist.sample()
            self.new_log_prob = dist.log_prob(self.new_action)

            # Calc loss
            log_prob = dist.log_prob(tf.squeeze(self.actions))
            loss = tf.reduce_mean(tf.squeeze(self.targets) - 0.2 * log_prob)
            self.train_op = optimizer.minimize(loss, var_list=_params('policy_net'))

    def get_action(self, s):
        action = self.sess.run(self.new_action, feed_dict={self.states: s[np.newaxis, :]})
        return action[0]

    def get_next_action(self, s):
        next_action, next_log_prob = self.sess.run([self.new_action, self.new_log_prob], feed_dict={self.states: s})
        return next_action.reshape((-1, 1)), next_log_prob.reshape((-1, 1))

    def update(self, s, a, target):
        self.sess.run(self.train_op, feed_dict={self.states: s, self.actions: a, self.targets: target})

更新功能

def soft_q_update(batch_size, frame_idx):
    gamma = 0.99
    alpha = 0.2

    state, action, reward, next_state, done = replay_buffer.sample(batch_size)
    action = action.reshape((-1, 1))
    reward = reward.reshape((-1, 1))
    done = done.reshape((-1, 1))

Q_target

v_ = value_net_target.get_value(next_state)
q_target = reward + (1 - done) * gamma * v_

V_target

next_action, next_log_prob = policy_net.get_next_action(state)
q1 = soft_q_net_1.get_q(state, next_action)
q2 = soft_q_net_2.get_q(state, next_action)
q = np.minimum(q1, q2)
v_target = q - alpha * next_log_prob

Policy_target

q1 = soft_q_net_1.get_q(state, action)
q2 = soft_q_net_2.get_q(state, action)
policy_target = np.minimum(q1, q2)

最佳答案

由于该算法对于离散策略和连续策略都是通用的，因此关键思想是我们需要可重新参数化的离散分布。然后，扩展只需通过更改策略分发类就可以从连续SAC中进行最少的代码修改。

有一种这样的分布-GumbelSoftmax分布。 PyTorch没有内置的功能，因此我仅从具有正确的rsample（）的近亲中扩展它，并添加正确的对数概率计算方法。借助计算重新设置参数的操作及其对数概率的能力，SAC可以用最少的额外代码很好地处理离散操作，如下所示。

    def calc_log_prob_action(self, action_pd, reparam=False):
        '''Calculate log_probs and actions with option to reparametrize from paper eq. 11'''
        samples = action_pd.rsample() if reparam else action_pd.sample()
        if self.body.is_discrete:  # this is straightforward using GumbelSoftmax
            actions = samples
            log_probs = action_pd.log_prob(actions)
        else:
            mus = samples
            actions = self.scale_action(torch.tanh(mus))
            # paper Appendix C. Enforcing Action Bounds for continuous actions
            log_probs = (action_pd.log_prob(mus) - torch.log(1 - actions.pow(2) + 1e-6).sum(1))
        return log_probs, actions


# ... for discrete action, GumbelSoftmax distribution

class GumbelSoftmax(distributions.RelaxedOneHotCategorical):
    '''
    A differentiable Categorical distribution using reparametrization trick with Gumbel-Softmax
    Explanation http://amid.fish/assets/gumbel.html
    NOTE: use this in place PyTorch's RelaxedOneHotCategorical distribution since its log_prob is not working right (returns positive values)
    Papers:
    [1] The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables (Maddison et al, 2017)
    [2] Categorical Reparametrization with Gumbel-Softmax (Jang et al, 2017)
    '''

    def sample(self, sample_shape=torch.Size()):
        '''Gumbel-softmax sampling. Note rsample is inherited from RelaxedOneHotCategorical'''
        u = torch.empty(self.logits.size(), device=self.logits.device, dtype=self.logits.dtype).uniform_(0, 1)
        noisy_logits = self.logits - torch.log(-torch.log(u))
        return torch.argmax(noisy_logits, dim=-1)

    def log_prob(self, value):
        '''value is one-hot or relaxed'''
        if value.shape != self.logits.shape:
            value = F.one_hot(value.long(), self.logits.shape[-1]).float()
            assert value.shape == self.logits.shape
        return - torch.sum(- value * F.log_softmax(self.logits, -1), -1)

这是LunarLander结果。 SAC学会了非常快速地解决它。



完整的实现代码位于SLM Lab的https://github.com/kengz/SLM-Lab/blob/master/slm_lab/agent/algorithm/sac.py

以下显示了Roboschool（连续）和LunarLander（离散）的SAC基准结果：https://github.com/kengz/SLM-Lab/pull/399