Scholarlink

<p>Low-Earth-orbit (LEO) mega-constellations turn satellite–ground scheduling into a large-scale, real-time combinatorial problem. This work presents a neural-combinatorial framework that couples policy-gradient reinforcement learning with a transformer-based pointer network. The overall solution is implemented as a modular pipeline composed of distinct components for preprocessing, training, scheduling, and emergency handling, wherein the learning component consists of a single policy-gradient agent. The policy operates on ephemeris-derived contact graphs where candidates are encoded by priority, start time, satellite ID, and ground-station ID; feasibility is enforced through masking and a reward that captures buffer, non-overlap, and cool-down constraints. In simulations on OneWeb (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\approx 650\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≈</mo> <mn>650</mn> </mrow> </math></EquationSource> </InlineEquation> sats) and Starlink (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\approx 3500\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≈</mo> <mn>3500</mn> </mrow> </math></EquationSource> </InlineEquation> sats), the method reduces the mean number of missed satellites by up to 54.3% versus a tuned greedy search and by up to 29.5% versus dynamic programming (DP) over 6–48&#xa0;h horizons, at the cost of modestly higher idle time (4.1–16.1%). End-to-end runtime is competitive at scale: for a 3500-satellite, 48&#xa0;h test the policy schedules in 366.2 s (DP 234&#xa0;s; greedy 410&#xa0;s). The approach learns feasibility, scales to constellation size, and adapts to changing task priorities without post-processing, offering a promising alternative to traditional planners.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automatic scheduling of satellite tracking tasks by means of policy-gradient reinforcement learning and transformer-based pointer networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月29日

查看全文
  • Ravi Dobariya,
  • Thomas Hobiger

摘要

Low-Earth-orbit (LEO) mega-constellations turn satellite–ground scheduling into a large-scale, real-time combinatorial problem. This work presents a neural-combinatorial framework that couples policy-gradient reinforcement learning with a transformer-based pointer network. The overall solution is implemented as a modular pipeline composed of distinct components for preprocessing, training, scheduling, and emergency handling, wherein the learning component consists of a single policy-gradient agent. The policy operates on ephemeris-derived contact graphs where candidates are encoded by priority, start time, satellite ID, and ground-station ID; feasibility is enforced through masking and a reward that captures buffer, non-overlap, and cool-down constraints. In simulations on OneWeb ( \(\approx 650\) ≈ 650 sats) and Starlink ( \(\approx 3500\) ≈ 3500 sats), the method reduces the mean number of missed satellites by up to 54.3% versus a tuned greedy search and by up to 29.5% versus dynamic programming (DP) over 6–48 h horizons, at the cost of modestly higher idle time (4.1–16.1%). End-to-end runtime is competitive at scale: for a 3500-satellite, 48 h test the policy schedules in 366.2 s (DP 234 s; greedy 410 s). The approach learns feasibility, scales to constellation size, and adapts to changing task priorities without post-processing, offering a promising alternative to traditional planners.