Su, C.G. Wang, Z.J. JACoW Publishing Geneva, Switzerland The reinforcement learning (RL) algorithm is utilized to control the low-energy beam transport (LEBT) and radiofrequency quadrupole (RFQ) in linear accelerators, with the aim of improving RFQ transmission efficiency, achieving high beam intensity, reducing debugging time, and improving operational efficiency. A neural network model is established as part of the Interaction environment to partially replace the Tracewin software for RL training proceess. The SAC algorithm is a reinforcement learning algorithm used to optimize control policies for continuous action spaces. By using the SAC algorithm and interacting with the neural network model, a policy was trained to control the LEBT solenoids, optimizing the RFQ transmission efficiency to above 95\% on the simulation software Tracewin. To test the generalization ability of the strategy, we applied it to a real accelerator and successfully validated its ability to optimize the RFQ transmission efficiency. The results demonstrate that RL policy trained in simulation-based environments can be applied on real accelerator control.