BiBHNet: A Bilateral Beam Hopping Neural Network for Joint Beam Pattern and Resource Allocation in GEO Satellite Communications

Abstract

Geostationary Earth orbit (GEO) high-throughput satellite (HTS) systems face the challenges of non-uniform traffic distributions and inter-beam interference under limited resources. Conventional beam hopping (BH) methods typically decouple multidimensional allocations or rely on iterative optimization because the mixed discrete-continuous action space grows rapidly with the number of beams and resource variables, which can limit allocation quality or increase online computation cost. Therefore, this paper proposes a single-agent deep reinforcement learning (DRL) framework based on proximal policy optimization (PPO), incorporating a bilateral BH neural network (BiBHNet) to jointly optimize time-space-frequency-power resources in an end-to-end manner under dynamic traffic. The finite-horizon BH process is formulated as a partially observable Markov decision process (POMDP). The policy generates one fixed-dimensional feasible resource action at each state transition, while queue evolution, discounted returns, and value estimation preserve inter-slot coupling. A bilateral-branch encoder integrates temporal and instantaneous features and supports parallel heterogeneous policy heads and a shared value head. Binary indexing encodes feasible beam patterns, whereas bandwidth and power parameters are sampled from Beta distributions and projected onto physical constraints. Training employs asynchronous trajectory collection and dual adaptive normalization of a multidimensional shaped reward. Simulation results demonstrate that BiBHNet effectively adapts to non-uniform traffic distributions while balancing throughput, delay, packet loss, and allocation fairness. On the evaluated platform, BiBHNet offers a trade-off between decision-generation latency and resource-allocation performance relative to the evaluated genetic-algorithm baseline.