Multi-Agent Meta Reinforcement Learning for Packet Routing in Dynamic Network Environments

Student: Shan Sun (University of California, Riverside)
Supervisor: Mariam Kiran (Lawrence Berkeley National Laboratory)

Abstract: Traffic optimization challenges, such as flow scheduling and completion time reducing, are difficult online decision-making problems in wide area networks. Previous works apply heuristics that rely on full knowledge of the system to design optimization algorithms. In this work, we explore building a model-free approach, applying multi-agent meta reinforcement learning to solve complex online control problem that generates optimal paths to reroute traffic. Focusing on decentralized solutions, our experiment aims to efficiently minimize the average packet completion time while reducing packet loss across complex network topologies. To evaluate, we test with a static topology and dynamically changing network topologies and compare results to the classical shorted path algorithm.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: PDF


Back to Poster Archive Listing