A Novel Reinforcement Learning-Based Multi-Objective Optimization Framework for Power Allocation in Multibeam Satellite Systems

Multibeam satellite systems (MSS) are becoming a key component of future 6G non-terrestrial networks, providing wide-area coverage and high-throughput connectivity. However, their performance critically depends on power allocation, which leads to a high-dimensional and non-convex multi-objective optimization problem: the system must simultaneously minimize unmet system capacity and total transmit power. To address this challenge, this paper benchmarks representative multi-objective optimizers for MSS resource allocation using a simulation-based system model as a unified evaluator. Under static traffic demands, we systematically compare NSGA-II, MOEA/D, and multi-objective particle swarm optimization (MOPSO). Experimental results show that, under the same computational budget, MOPSO outperforms NSGA-II and MOEA/D in terms of hypervolume and inverted generational distance, establishing MOPSO as a strong baseline for MSS power allocation. Building on this baseline, we propose a reinforcement learning-basedMOPSO(RL-MOPSO), where a lightweight policy adaptively tunes swarm parameters based on online population statistics. We then evaluate RL-MOPSO in dynamic scenarios that emulate realistic channel conditions and demand variations. In these dynamic scenarios, RL-MOPSO achieves significant additional gains: it improves hypervolume by about 10%, reduces average USC by 40%, and lowers average transmit power by more than 90% compared with MOPSO. These results demonstrate that integrating multi-objective swarm optimization with reinforcement learning provides an effective approach for multibeam satellite systems power allocation.