78f46822aec6cea5e216e1f7404d351835c54346,yarll/agents/sac.py,SAC,train,#SAC#,92

Before Change

            values = self.value_network(state0_batch)
            value_loss = tf.reduce_mean(tf.square(values - value_target))
        advantage = tf.stop_gradient(action_logprob - new_softq + values)
        with tf.GradientTape() as actor_tape:
            _, action_logprob = self.actor_network(state0_batch)
            actor_loss = tf.reduce_mean(action_logprob * advantage)

After Change

        with tf.GradientTape() as softq_tape:
            softq = self.softq_network(state0_batch, action_batch)
            softq_loss = tf.reduce_mean(tf.square(softq - softq_targets))
        with tf.GradientTape() as value_tape:
            values = self.value_network(state0_batch)
        with tf.GradientTape() as actor_tape:
            actions, action_logprob = self.actor_network(state0_batch)

Instances