RAMA reduces latency in LEO satellite networks with retrieval-augmented multi-agent DRL
Original title: RAMA: Retrieval-Augmented Multi-Agent Routing with Episodic Memory for LEO Satellite Networks
RAMA integrates episodic memory into multi-agent DRL to cut mean end-to-end latency by 13.4% and 99th-percentile latency by 66.7% in a 140-satellite Kepler constellation.
The paper proposes RAMA, a retrieval-augmented multi-agent deep reinforcement learning framework that equips each satellite with local observations, routing history, neighboring state data, and retrieved episodic memory entries. A Transformer-based decision network trained via Approximate Information State auxiliary objectives enables effective use of past experiences. Evaluations on a 140-satellite Kepler constellation show mean latency reductions of 13.4% and 99th-percentile reductions of 66.7% versus double deep Q-network routing, with RAMA remaining stable during online updates while DDQN degrades. On the 1,584-satellite Starlink shell 1, mean latency drops by 93.2% and online updates yield an additional 25.5% improvement.
- 13.4% mean latency reduction in 140-satellite Kepler constellation
- 66.7% 99th-percentile latency reduction in Kepler setup
- RAMa stable during online updates while DDQN degrades
- 93.2% mean latency reduction in 1,584-satellite Starlink shell 1
- further 25.5% mean latency drop from online updates in Starlink simulations
- Tags
- leo-satellitesmulti-agent-drlretrieval-augmentedepisodic-memorysatellite-networksreinforcement-learning
- Products
- Kepler constellationStarlink
From the original article
1 min read at arxiv.orgLow-Earth orbit (LEO) satellite networks present a challenging routing environment due to rapidly changing topology, traffic, and link conditions. Existing deep reinforcement learning (DRL) routing methods typically rely on current observations and information encoded in network weights, limiting their ability to exploit successful past experiences.
By Xin Ming Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.