An In-Depth Analysis of a Machine-Learning-Based Network Routing Protocol For Networking in a Highly Mobile Topology
NAVAL POSTGRADUATE SCHOOL MONTEREY CA MONTEREY United States
Pagination or Media Count:
This thesis studies the performance of a machine-learning-based DTN routing protocol, QGeo. QGeo is based on the reinforcement learning model called Q-learning whereby an agent in some context takes an action, gains a reward and adapts its decision-making policy based on the rewards value. QGeo is implemented in the ns-3 simulator, and the implementation in this work is based on the previously implemented GAPR protocols. QGeo is then tested in ns-3 alongside GAPR, GAPR2 and GAPR2a, as well as the more commonly known Epidemic, Vector and Centroid DTN protocols. Testing is performed rigorously across four simulation scenarios. The Helsinki scenario simulates mobile traffic in a city, the Omaha and Bold Alligator scenarios simulate amphibious military exercises with various properties, and the Swarm scenario simulates the behavior of a drone swarm based on real-world sensor flight data. This thesis ultimately shows that QGeo is a highly selective protocol in terms of making forwarding decisions, based primarily in the Q-learning mechanism. This thesis also advances the research previously done at the Naval Postgraduate School in DTN research and development by furthering the testing effort of the protocols that have been implemented. Finally, an added benefit of this study is the incorporation of the Swarm scenario to the DTN testbed, increasing the range of testing capability for comparison of DTN routing protocol characteristics.
- Computer Systems