loss function for the policy gradient is:
self.g_loss = -tf.reduce_sum( tf.reduce_sum( tf.one_hot(tf.to_int32(tf.reshape(self.x, [-1])), self.num_emb, 1.0, 0.0) * tf.log( tf.clip_by_value(tf.reshape(self.g_predictions, [-1, self.num_emb]), 1e-20, 1.0) ), 1) * tf.reshape(self.rewards, [-1])
And I can hardly understand how the g_loss term is the same as the objective function in the paper, and also the relationship with the policy gradient approximation in the paper.
loss function for the policy gradient is:
self.g_loss = -tf.reduce_sum( tf.reduce_sum( tf.one_hot(tf.to_int32(tf.reshape(self.x, [-1])), self.num_emb, 1.0, 0.0) * tf.log( tf.clip_by_value(tf.reshape(self.g_predictions, [-1, self.num_emb]), 1e-20, 1.0) ), 1) * tf.reshape(self.rewards, [-1])
And I can hardly understand how the g_loss term is the same as the objective function in the paper, and also the relationship with the policy gradient approximation in the paper.