An Entropic Value-at-Risk-Inspired Load-Balancing Routing Algorithm for LEO Satellite Networks

In sixth-generation (6G) space–air–ground integrated networks, risk-neutral reinforcement-learning-based routing methods for low Earth orbit (LEO) satellite networks remain susceptible to severe tail latency. To address this issue, this paper proposes a novel Entropic Value-at-Risk (EVaR)-inspired routing algorithm with risk regularization for low-return actions. The routing process in an LEO satellite network is formulated as an event-driven partially observable Markov decision process. Within a multi-agent soft actor–critic framework, an EVaR-inspired state-level risk proxy for low-return actions is introduced and incorporated into the actor loss as a risk regularizer. This design enables the joint optimization of expected routing returns and the mitigation of low-return action risk. Simulation results obtained in a simplified LEO network using Iridium orbital parameters demonstrate that, compared with the multi-agent soft actor–critic algorithm, the proposed algorithm reduces the average end-to-end delay and the 99th-percentile end-to-end delay by 19.2% and 32.9%, respectively. It also reduces the average and maximum queueing delays by 79.4% and 57.0%, respectively, with approximately no packet loss observed during the simulations. These results demonstrate that the proposed algorithm enables efficient and reliable routing in LEO satellite networks under high-load conditions.