Dueling DQN splits the action-value function into a state value head V(s) and an advantage head A(s, a), then aggregates them back into Q(s, a).
