With the rapid development of technology, people's demand for wireless spectrum is increasing. However, due to limited spectrum resources, how to effectively utilize these resources has become a major challenge in the field of radio. To address this is‐ sue, we establish a cognitive wireless network model where primary and secondary users share the same spectrum resources and work in a non cooperative manner to improve the throughput of secondary users. Then, we use the SumTree Sampling Deep QNetwork (ST-DQN) algorithm for power control to ensure priority and diversity in sample selection. Finally, a series of simulation experiments are conducted using Python to compare and analyze the performance indicators of reward, loss function, and sub user throughput with traditional Q-learning and free exploration algorithms. We find that the ST-DQN algorithm performs better in power control.