Adaptive Modulation Using Multi-Objective Reinforcement Learning for LEO Satellites

FELIPE PASQUEVICH; GRACIELA CORRAL BRIONES

Cleveland, Ohio.

Workshop; Cognitive Communications for Aerospace Applications (CCAA) Workshop 2021; 2021

IEEE Cleveland Section and NASA Glenn Research Center

In this paper, an emerging Machine Learning technique, named Multi-Objective Reinforcement Learning (MORL), is applied and analyzed aiming to achieve a two-fold optimization in a Satellite-To-Ground communication. The objectives pursued in this work using MORL are to minimize the Bit Error Rate and keep simultaneously the best performance in terms of the maximum bit rate transmission. The scenario under test consists of a Low-Earth Orbit satellite moving in a circular orbit while establishing a line-of-sight communication with a Ground Station. Two approaches are evaluated, the Weighted Sum and the Thresholded Lexicographic Q-Learning. We show that these two approaches can not find all the solutions. To alleviate this situation, a novel proposal is considered based on an inverse scalarization function that allows to select any solution from the Pareto front. We show that the proposed algorithm is able to implement a suitable operation for many different digital modulation schemes to obtain the maximum throughput.