LEO SATELLITE BASED QUANTUM INTERNET: DEEP REINFORCEMENT LEARNING FOR OPTIMAL RESOURCE ALLOCATION UNDER REQUEST CONSTRAINTS

This thesis addresses the development of a global infrastructure for the Quantum Internet. Since terrestrial optical fibers suffer from high signal attenuation over long distances during qubit transmission , the explored solution leverages LEO satellite constellations to distribute entangled photons between widely separated ground stations. However, managing a spatial network of this kind is highly complex as satellites move at extremely high speeds, continuously altering the network’s topology and geometry. Added to this is the inherent fragility of quantum information; photons lose fidelity due to background noise and atmospheric turbulence, and quantum memories also experience rapid decoherence. For these reasons, extremely strict time constraints are imposed. Faced with this extreme dynamism, classical Greedy routing algorithms prove inadequate in managing resources, minimizing network latency, and handling longterm queues. To overcome this obstacle, an Artificial Intelligence-based approach was introduced by implementing a Deep Reinforcement Learning algorithm called Proximal Policy Optimization (PPO). This algorithm was trained to make optimal scheduling decisions aimed at maximizing the number of successfully satisfied requests, where each request strictly requires the generation of three entangled photon pairs with a fidelity above 90% and minimizing connections that expire due to time limits. The obtained results demonstrate that the PPO agent achieves significantly better performance in terms of throughput, about the 50% more (i.e., satisfied requests per second) and in reducing the number of expired requests compared to the deterministic Greedy approach.