An analysis of Mirror-Descent Soft Actor-Critic
This work analyzes Soft Actor-Critic (SAC) when using a policy mirror descent target instead of the standard Gibbs target. It proves that by tuning the mirror descent step size ($\lambda$), one can directly control the tracking error, establishing an explicit trade-off between policy improvement and tracking accuracy. It also shows that the classical Gibbs target introduces a persistent, non-vanishing error term.