Research On The Preparation Method Of Quantum State Based On Reinforcement Learning

The control design of quantum systems is considered to be a necessary requirement for establishing powerful quantum information technology,but the experiments of quantum systems control often face different constraints,mainly including the minimum control quantity(only required to be bounded or the control form is unlimited)and the realization of control in the shortest time.Quantum state preparation is a key task in the field of quantum system control.In this paper,reinforcement learning algorithm is used to solve the problem of high-fidelity quantum state preparation under limited conditions.The main work is as follows:(1)A quantum state preparation method based on reinforcement learning is proposed.The Q value evaluation network performance of existing quantum state preparation methods based on deep reinforcement learning is poor,and network parameters are updated in the form of random sampling,resulting in a large number of quantum gates and low fidelity required for quantum state preparation under limited conditions.To address the above problems,a novel quantum state value evaluation network is designed to predict the optimal Q value after performing quantum operations and reduce the number of quantum gates required.In addition,the parameter updating method based on sample priority is adopted to accelerate the convergence speed of the algorithm and quickly learn the optimal control strategy,thus improving the fidelity of the final quantum state.The Q-learning,Policy Gradient and Deep Qlearning algorithm are used to verify the effectiveness in simulation experiments.Compared with other quantum state preparation methods based on reinforcement learning,our method have different degrees of improvement in convergence speed and the number of training episodes required when the fidelity of quantum states is closed to 1.(2)A quantum state preparation method based on difference-driven reinforcement learning is proposed.The two-qubit system has a large state space,and most of the existing quantum state preparation methods use step reward function,which results in slow convergence speed and difficulty to prepare the desired target quantum state with high fidelity under limited conditions.To address the above problems,a weighted differential dynamic reward function is designed for the quantum state preparation task of the two-qubit system to help quickly obtain the maximum expected cumulative reward.In addition,the adaptive ?-greedy action selection strategy is adopted to achieve the balance of exploration and utilization to a certain extent,thus improving the fidelity of the final quantum state.The simulation results show that the proposed algorithm can prepare quantum state with high fidelity under limited conditions.Compared with other algorithms,it has different degrees of improvement in the convergence speed and the fidelity of the final quantum state.