Deep reinforcement with spectrum series learning control for a mode-locked fiber laser

Abstract

A spectrum series learning-based model is presented for mode-locked fiber laser state searching and switching. The mode-locked operation search policy is obtained by our proposed algorithm that combines deep reinforcement learning and long short-term memory networks. Numerical simulations show that the dynamic features of the laser cavity can be obtained from spectrum series. Compared with the traditional evolutionary search algorithm that only uses the current state, this model greatly improves the efficiency of the mode-locked search. The switch of the mode-locked state is realized by a predictive neural network that controls the pump power. In the experiments, the proposed algorithm uses an average of only 690 ms to obtain a stable mode-locked state, which is one order of magnitude less than that of the traditional method. The maximum number of search steps in the algorithm is 47 in the 16°C–30°C temperature environment. The pump power prediction error is less than 2 mW, which ensures precise laser locking on multiple operating states. This proposed technique paves the way for a variety of optical systems that require fast and robust control.