b7dbbfbf4111698145bb9e0bf2e34e36fef0430c,rllib/agents/sac/sac_tf_policy.py,,sac_actor_critic_loss,#Any#Any#Any#Any#,213

Before Change

            y_true=q_t_selected_target, y_pred=q_t_selected)
    ]
    if policy.config["twin_q"]:
        critic_loss.append(0.5 * tf.keras.losses.MSE(
            y_true=q_t_selected_target, y_pred=twin_q_t_selected))

    // Alpha- and actor losses.
    // Note: In the papers, alpha is used directly, here we take the log.
    // Discrete case: Multiply the action probs as weights with the original

After Change

        td_error = base_td_error

    // Calculate one or two critic losses (2 in the twin_q case).
    prio_weights = tf.cast(train_batch[PRIO_WEIGHTS], tf.float32)
    critic_loss = [tf.reduce_mean(prio_weights * huber_loss(base_td_error))]
    if policy.config["twin_q"]:
        critic_loss.append(
            tf.reduce_mean(prio_weights * huber_loss(twin_td_error)))

Instances