Reinforcement Learning strategies for Quantum Optimization Algorithms

We present a strategy based on classical control of Quantum devices using Reinforcement Learning. Our strategy is applied to Quantum algorithms designed for classical optimization problems such as the QAOA and Quantum Annealing. Our method provides optimal control of the Quantum device following a reformulation the Quantum algorithm as an environment where an autonomous classical agent interacts and performs actions to achieve higher rewards. This formulation allows a hybrid classical-Quantum device to train itself from previous executions using both model-based and model-free Reinforcement Learning to control the degrees of freedom of the Quantum Algorithm. Our approach makes a selective use of Quantum measurements to complete the observations of the Quantum state available to the agent. We run tests of this approach on several classical optimization problems, obtaining optimal results for problem instances with N > 20 decision variables. We show how this formulation can be used in variational algorithms to transfer the knowledge from shorter training episodes to reach larger circuit depths and deliver better results.