Avaliação experimental do uso de Aprendizado por Reforço para Alocação Dinâmica de Canais em Cenários LEO Multifeixe Simplificados

Abstract

Low Earth Orbit (LEO) satellite networks emerge as a relevant alternative to expand connectivity in remote regions, hard-to-reach areas, and scenarios with limited terrestrial infrastructure. However, these networks face challenges related to satellite mobility, limited radio resources, and uneven demand distribution. In this context, this work experimentally evaluates the use of reinforcement learning for dynamic channel allocation in simplified multibeam LEO scenarios, comparing classical heuristics with an approach based on Proximal Policy Optimization (PPO). As a contribution, an integrated experimental environment was developed, combining a Python decision simulator with a complementary ns-3 simulation, enabling the analysis of allocation policies both in the decision-making process and in network-level metrics, such as throughput, blocking, and delay. The results indicate that adaptive policies outperform uniform or rotational strategies in scenarios with asymmetric demand. The Greedy Backlog heuristic achieved the highest served traffic and the lowest blocking rate, while the PPO agent achieved the lowest average delay in the main scenario and performance close to intermediate heuristics. Tests under different traffic load conditions, hotspot demand concentration, and channel availability showed that PPO maintained consistent behavior in throughput, blocking, and delay metrics. Therefore, the results indicate that reinforcement learning has potential for adapting to demand changes and resource distribution in simplified multibeam LEO scenarios.