Determining optimal policies for quantum networks – A Reinforcement Learning approach*

Near-term implementation of quantum networks must overcome current hardware limitations such as link losses, non-ideal measurements, and low decoherence-time quantum memories. In this context, it has been shown that optimizing figures of merit such as average connection time, largest connected cluster size, and average network fidelity, can be formulated as a decision process. An optimal protocol or policy (series of decisions) can therefore be determined using reinforcement learning (RL), which is well suited to handle decision processes. In the present work, we simulate a near-term quantum network based on entanglement distribution between nodes and optimize the figures of merit using RL. More specifically, we use a model-independent algorithm called Q-learning. Nodes act as agents and perform actions to establish entanglement with other nodes. Nodes can collaborate to achieve optimal values for the figures of merit. Previous theoretical work has suggested that policies must depend on hardware limitations. We parametrize these limitations in a platform-independent way and show that optimal policies that Q-learning yields depend on these parameters. We then compare our findings with the limitations posed by the theory. We also study the role of local versus global collaboration between nodes, and how this effects the policies and network outcomes. To explore the applicability of the RL framework for policy optimization, we implement different network architectures and multipartite entanglement generation between nodes.