Abstract: Stable operation of lasers requires a constant current source control system with high precision and stable output current. An algorithm based on the combination of central differential Kalman filtering （CDKF） and improved radial basis function （RBF） neural network adaptive PID control， named CDKF-RBFPID， is proposed to address the problem that the output accuracy of laser constant current source system is poor in noisy environment and the parameters are difficult to be adjusted by PID algorithm. By using CDKF， we update the state and covariance of the constant current source system， so as to filter out the state noise and measurement noise in the system. To accomplish adaptive parameter tuning， the RBF-PID parameters are modified using the reinforcement learning Actor-Critic architecture. Comparing the output current of the constant current source system and the output power of a laser， the experimental results demonstrate that the CDKF-RBFPID method can effectively lessen the impact of noise on the system， further enhance the accuracy of the constant current source output current and the stability of the laser output power， with the response time improved by 58.3%， the steady-state error reduced by 71.4%， and the output current control accuracy reaching 1%.

Key words: constant current source, PID control, CDKF, RBF neural network, Actor-Critic