Beyond Performance: Explaining Non-Intuitive Deep Reinforcement Learning Actions in Elastic Optical Networks
Paper in proceeding, 2025
We develop a Deep Reinforcement Learning (DRL) agent for the RMSA problem and improve the Shapley Value for Explaining Reinforcement Learning (SVERL) explainability framework by integrating policy sensitivity and feature interdependence for the RMSA problem. We then explain the proactive rejection of lightpath requests.
Optical networks
Shapley Value for Explaining Reinforcement Learning (SVERL)
Explainable AI
Author
Louma Mehyeddine
Lebanese University
Carlos Natalino Da Silva
Chalmers, Electrical Engineering, Communication, Antennas and Optical Networks
Alaa Amro
Lebanese University
Jean Pierre Asdikian
Polytechnic University of Milan
Ihab Sbeity
Lebanese University
Guido Maier
Polytechnic University of Milan
Paolo Monti
Chalmers, Electrical Engineering, Communication, Antennas and Optical Networks
Sebastian Troia
Polytechnic University of Milan
Omran Ayoub
University of Applied Sciences and Arts of Italian Switzerland (SUPSI)
Proceedings of the 51st European Conference on Optical Communication, ECOC 2025
51st European Conference on Optical Communication
Copenhagen, Denmark,
Copenhagen, Denmark,
Efficient Confluent Edge Networks (ECO-eNET)
European Commission (EC) (EC/HE/101139133), 2024-01-01 -- 2028-12-31.
Areas of Advance
Information and Communication Technology
Subject Categories (SSIF 2025)
Communication Systems
Computer Sciences