Erol Gelenbe - The Cognitive Packet Network - Reinforcement based Network Routing with Random Neural Networks

Organisé par : 
L'équipe des Keynote Speeches : Nicolas Peltier, Renaud Lachaize, Jérôme David
Intervenant : 
Erol Gelenbe (Imperial College London)


Erol Gelenbe is the Dennis Gabor Professor of Electrical and Electronic Engineering at Imperial College London. Gelenbe invented trailblazing mathematical models including G (Gelenbe)-Network and Random Neural Network (RNN) models that allow for performance evaluation and analysis of computer systems and networks. Gelenbe’s fundamental contributions in these areas have also been instrumental in allowing networks to operate seamlessly without overloading. Along with colleagues, he is also credited with inventing an early computer architecture that allowed voices and images to travel over multi-hop and multi-path computer and communications networks. Gelenbe’s most recent research interests include Software Defined Networks (SDNs), energy savings in information and computing technology (ICT), security in networks, and reinforcement and deep learning within neural networks. Gelenbe is a Fellow of ACM, IEEE and the Institution of Engineering and Technology (IET) UK. He was also elected a Fellow of the National Academy of Technologies of France, and the Science Academies of Belgium, Hungary, Poland and Turkey. In 2017, Gelenbe received the Mustafa Prize, a $500,000 biennial science and technology award which aims to rival the Nobel prizes, that is bestowed by the Mustafa Foundation of Iran. He is also a recipient of the 2008 ACM SIGMETRICS Achievement Award, given annually to an individual who has made long-lasting influential contributions to the analysis and evaluation of computer and communication system performance, and several other awards, including the Grand Prix France Telecom 1996 of the French Academy of Sciences.


Réalisation technique : Antoine Orlandi | Tous droits réservés

The Cognitive Packet Network (CPN) is an experimental network routing protocol which uses specific Quality of Service (QoS) objectives incorporated in a Goal Function, together with network measurement by Smart Packets (SPs). It updates neural network based Oracles in routers using Reinforcement Learning, in order to dynamically select network paths so that end users can convey their payload traffic with a performance that matches the Goal as closely as possible. The Goal can include conventional QoS metrics such as delay and loss, as well as Real-Time objectives, as well as newer metrics of interest including Energy Consumption and Security. Payload traffic is forwarded using source or segment routing, selected through the reinforcement learning approach, while SPs conduct their exploration using a node by node process by seeking the best direction from each Oracle. CPN has been implemented in various contexts: on 10-40 node test-beds, on an intercontinental scale as an overlay network, within SDN routers, and as a means to convey task requests over the Internet to Cloud servers. Our presentation will detail the CPN algorithm and the Random Neural Networks that are used to implement the Oracles. We will also present relate experimental measurements and results. The work has appeared in a variety of journals and conferences including CACM, Proceedings IEEE, IEEE J. Sel. Areas in Comms.