Abstract: Aiming at the problem that the conventional soft actor-critic (SAC) algorithm lacked exploration capability and state representation in complex environments, we proposed an improved soft actor-critic (ISAC) algorithm. Firstly, the ISAC algorithm  introduced a prioritized experience replay (PER) mechanism, which dynamically evaluated the priority of experience samples by using the temporal differential errors, thereby enhancing the utilization of crucial experiences and improving learning efficiency of the algorithm. Secondly, the algorithm integrated  generative Transformer architecture  into the actor network to strengthen its ability to dynamically capture state features, thereby significantly improving its performance in complex optimization tasks. Finally, we conducted an application experiment  on the dynamic scheduling optimization problem of university logistics staff. The experimental results show that, compared with the original  SAC algorithm and the classic deep Q-network (DQN) algorithm, the proposed ISAC algorithm has smaller errors in dynamically fitting human resource demand, which effectively demonstrates its 
advantages and practicality in practical applications.

Key words: deep reinforcement learning, soft actor-critic algorithm, prioritized experience replay, Transformer architecture,  , logistics management