Project Coordinator (EU) :Rheinisch-Westfälische Technische Hochschule Aachen
Country of the EU Coordinator :Germany
Organisation Type :Academia
Project participants :
Haris Kremo (RWTH) - M.S. and Ph.D. degrees in electrical and computer engineering from Rutgers, the State University of New Jersey, USA. His focus is on experimenting and prototyping on physical and medium access layers of the protocol stack. He is the principal investigator overseeing the administrative and scientific aspects of the project. He will also engage in building of the experimental setup, and other practical aspects related to execution of experiments.
Ivan Šeškar (Rutgers University,Ivan Šeškar (Rutgers University) - Chief Technologist at WINLAB, Rutgers University responsible for experimental systems and prototyping projects. He is currently the program director for the COSMOS project responsible for the New York City NSF PAWR deployment, the PI for the NSF GENI Wireless project, which resulted in campus deployments of LTE/WiMAX base stations at several US universities, and the PI for the NSF CloudLab deployment at Rutgers. Ivan is a co-chair of the IEEE Future Networks Testbed Working Group, a Senior Member of the IEEE, a member of ACM and the co-founder and CTO of Upside Wireless Inc. He acts as a co-principal investigator within the US side of the consortium with know-how on COSMOS testbed infrastructure and implementation.
Kenan Turbić (RWTH) - M.Sc. degree from the University of Sarajevo in 2011, and a Ph.D. degree (Hons.) in Electrical and Computer Engineering from IST, University of Lisbon, in 2019. In 2019 and 2020 he was a postdoctoral researcher at the INESC-ID research institute, Lisbon, Portugal. Currently he is a postdoctoral researcher and a teaching assistant at the Chair for Distributed Signal Processing, ICE, RWTH Aachen, Aachen, Germany. His main research interests are wireless channel modeling and estimation, signal processing for wireless communications, and application of machine learning methods for improved system performance. He will contribute to the algorithmic development and its deployment.
Amna Kopić (RWTH) - received a B.Sc. and an M.Sc. degree in Electrical Engineering from the Faculty of Electrical Engineering, University of Sarajevo, Bosnia and Herzegovina, in 2017 and 2019, respectively. While with the University of Sarajevo, she received the Silver and the Golden Badge awards for academic excellence. Currently, she is a research assistant at RWTH Aachen University, where she is working towards a doctoral degree. Her research interests include the areas of digital signal processing and machine learning, with applications related to wireless communication systems. Her role in the project is to implement and evaluate different ML techniques and conduct experiments.
State of US partner :New Jersey
Starting date :
Experimental Study of Context Based Routing Using Deep Reinforcement Learning
Our experiments will be conducted on the COSMOS testbed. The key NGI technologies we consider are artificial intelligence (AI) and 5G radio access. The best match among the NGI topics is “Discovery and Identification Technologies”. Relevant vertical sectors are telecommunication industry, automotive, content distribution, urban management and IoT.
We consider two sources of contextual information. One is an external bird’s-eye view camera, which creates periodic images of its surroundings (urban intersection) every few seconds. These images are analyzed for the presence of mobile or stationary entities, like vehicles or people, to deduce the changes in demand for network services. The second internal source of contextual information comes from the analysis of data streams running through the routers. The advantage of the combination of these two contextual information sources is in the fact that analysis at the routers is reactive.
Even fast reinforcement learning is reactive in the sense that the underlying algorithm must learn how to recognize and then recognize fluctuations in the network traffic when they occur. From a visual inspection of the geographic area with high density of diverse network nodes, one can build inference and anticipate issues as potential bottlenecks in radio access networks, type of network load (large throughput, short latency…), or exceptional events such as the emergency situations.
The external contextual information can, for instance, include examples such as recognition of automated vehicles with time-critical communication demands, pedestrians equipped with smartphones, buses full of IoT devices and passengers with their own wireless devices, or estimation of the number of customers entering/leaving a store.
In general, the envisioned experiments provide a test ground to employ different ML algorithms and different contextual information sources, and to mix and match them together to explore practical features and limitations of such synergy.
Implementation plan :
The full project development is structured in four phases:
- Analysis of suitable ML techniques;
- Selection of promising contextual information sources.
- on a sandbox testbed at RWTH;
- on the COSMOS testbed.
- Sandbox testbed experiments, in parallel with the implementation phase;
- Full-scale experiment on the COSMOS testbed.
Analysis and reporting
- Analysis of experimental results.
- Dissemination of the project findings.
- Sharing the code and gathered data on publicly available repositories.
The initial phase includes evaluation of various ML methods in order to select the most ideal and appropriate one for the problem at hand, where reinforcement learning neural network architectures present themselves as promising candidates. For validated and high-performance implementation of these ML algorithms, we will rely on well-known ML software libraries, such as PyTorch and Keras/Tensorflow. The selected algorithm will be deployed and tested in a sandbox testbed at RWTH, where only a few representative nodes are implemented.
After the prototype at RWTH is successfully tested, the system will be deployed in the COSMOS testbed, with its the user tier located in West Harlem, New York. The computing resources dedicated to the emulation of edge servers are placed on the Columbia University campus. Core servers at WINLAB facility in North Brunswick, New Jersey, will be used to emulate a data center. Inside of the testbed perimeter we will deploy a few testbed nodes on vehicles, some of the data sources will be carried around by pedestrians, and a fraction them will be stationary. Data traffic generated by the nodes will be different in terms of throughput and latency, and it will be changing over the course of experiments.
Impact 1: Enhanced EU – US cooperation in Next Generation Internet, including policy cooperation.
Impact 2: Reinforced collaboration and increased synergies between the Next Generation Internet and the Tomorrow's Internet programmes.
Impact 3: Developing interoperable solutions and joint demonstrators, contributions to standards.
Impact 4: An EU - US ecosystem of top researchers, hi-tech start-ups / SMEs and Internet-related communities collaborating on the evolution of the Internet.
The main outcome expected from this project is a prototype of a smart heterogeneous network that achieves:
- Better KPIs in comparison to the industry standard, with real-time or near real-time packet routing between the data sources and the testbed cloud.
- Faster convergence to optimal system configurations when the context information is included.
- Faster reaction to dynamic changes in the distribution and characteristics of the input traffic.
The developed system prototype will contribute to a better understanding of the potential improvements in the system performance by exploitation of context information.
The project will evolve according to timeline scheme below:
The following milestones are indicated in the timeline:
- M1: Completed evaluation of the candidate ML algorithms, enhanced with heterogeneous contextual information.
- M2: Finalized implementation of the proposed solution on the sandbox testbed at RWTH.
- M3: Completed analysis of experimental data and formalized deliverables.
The following deliverables will be produced according to the indicated schedule:
- D1: Conference paper or demo displaying the experimental setup.
- D2: The code and experimental traces for the research community.
- D3: Journal paper describing technical developments and the main findings.
Future Plan :
The intention with these experiments is to address TRL 3: Research to Prove Feasibility and open doors to move the proposed concept toward TRL 4: Technology Development. The most important evaluation parameter will be the relative reduction in latency for the time-critical data, achieved by the proposed algorithms. The results will be compared against the case with fixed data priority.