discrete_stochastic_mlp.py 文件源码

def sample_action(self, observation):
        """
        Samples an action from \pi_\theta(a|s)

        tf ops are eliminated on purpose here since this is a hot code path and
        were optimizing for CPU usage...or maybe tf.multinomial is just slow in general.

        Using TF ops:
        sample_action_op = tf.squeeze(tf.nn.softmax(self.net.logits))
        action = tf.multinomial(sample_action_op)
        """

        # TODO: ensure this works when num_actions > 1
        action_probs = self.net.sess.run(
            self.net.logits,
            {self.net.obs: [observation]}
        )[0]
        action = np.random.choice(np.arange(len(action_probs)), p = self.softmax(action_probs))

        return [action]