ECE Faculty Dario Pompili received a new supplement award from NSF
This Supplement is awarded by the National Radio Dynamic Zones (NRDZ) program and will focus on NextG Spectrum Sharing for O-RAN-based Emergency Networking via Deep Multi-Agent Reinforcement Learning.
Overview:
The Next radio Generation (NextG) of mobile networks, 6G and beyond, is expected to speed up the transition from monolithic and inflexible networks to agile and distributed networking elements that rely on “virtualization”, “softwarization”, openness, intelligent and yet fully interoperable Radio Access Network (RAN) components. Within the current virtual RAN (vRAN) and cloud RAN (C-RAN) concepts, spectrum sharing offers a natural solution to provide more efficient utilization of the spectrum by enabling different users or systems to share the same frequency bands dynamically. However, the traditional RAN framework involves large vendors creating proprietary hardware and software. As a result, in 2018, the O-RAN alliance was formed to realize the NextG cellular networks with flexible multi-vendor network infrastructure to the telecom operators. By deploying O-RAN, the network operators can significantly reduce the operational cost in a dense environment compared to vRAN and C-RAN. Due to the dynamic nature of spectrum policies and the unpredictability of unlicensed usage, spectrum sharing will require RANs to change their operational parameters intelligently and according to the current spectrum context. Although existing RANs do not allow for real-time reconfiguration, the fast-paced rise of the open RAN movement and of the O-RAN framework for NextG networks, where the hardware and software portions of the RAN are logically disaggregated, will allow seamless reconfiguration and optimization of the radio components.
Intellectual Merit:
In light of the above, a Multi-Agent Reinforcement Learning (MARL)-based cross-layer communication framework, “RescueNet”, was developed by the PI for self-adaptation of nodes in emergency networks. The use of mission policies, which vary over time and space, enabled graceful degradation in the Quality of Service (QoS) of the incumbent networks (only when necessary) based on mission policy specifications. Specifically, RescueNet was designed to solve the spectrum-sharing problem in emergency networking; however, this framework has limitations. Firstly, it converges slowly and requires a large amount of data to train, indicating that the RL agents need to interact with the environment extensively, which is time and bandwidth consuming. Secondly, RescueNet adopts Q-learning as the basic policy. Nonetheless, Q-learning cannot handle continuous action spaces, which poses additional constraints to the applications that RescueNet can handle.
To deal with these limitations, as part of this Supplement Award, the PI will perform two research tasks. In Task 1, the PI will focus on adopting MARL for emergency networking in O-RAN architecture. Specifically, in Task 1.A, the PI proposes knowledge sharing among agents to aid RescueNet in terms of convergence rate. Experienced agents will transfer their knowledge to new agents so that the new agents can avoid starting from scratch. In Task 1.B, the PI proposes to combine RescueNet with O-RAN leveraging its Near-Real-Time (NRT) Radio Intelligent Controller (RIC), and proposes intelligent spectrum management unit and policy decision unit for NRT decision making. Then, in Task 2, the PI will propose deep Hierarchical Multi-Agent Actor-Critic (HMAAC) by adding a high-level policy for coordination to the Multi-Agent Actor Critic (MAAC) framework to tackle the convergence rate and action space problems. With the help of deep neural networks and hierarchical structures, coordination and knowledge sharing will become more efficient. In Task 2.A, the PI will propose decentralized HMAAC, where the high-level policy is assigned to multiple agents to avoid single-point-failure. Finally, in Task 2.B, the PI will propose to extend HMAAC to the multi-team scenario where multiple teams can cooperate together. In this setting, another hierarchy will be added to HMAAC to coordinate among teams; furthermore, a communication reward will be included to encourage the agents to maintain communication.
More info can be found here: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2030101&HistoricalAwards=false
REFAI - "Ultra Low-Power Machine Learning via Hardware and Software Co-design"
Guest Lecture - Prof. Wujie Wen, North Caroline State University
You're invited to this Rutgers Efficient AI (REFAI) Seminar
Yingying Chen's group wins Best Paper Runner-Up Award at IEEE MASS 2023
Yingying Chen's group won the Best Paper Runner-Up Award for their paper entitled "Towards Efficient Privacy-Preserving Top-k Trajectory Similarity Query" at IEEE MASS 2023. The 20th IEEE International Conference on Mobile Ad-Hoc and Smart Systems (MASS 2023) is the premier conference for mobile ad-hoc networks and smart systems, defined broadly. As wireless ad-hoc networks continue to evolve and specialize into a number of application scenarios and environments, and sensor-based systems and technologies increasingly permeate our everyday life and become the inner fabric of the Internet of Things and cyber-physical systems, the unfolding of smart environments such as smart cities, smart farming, smart healthcare, and smart manufacturing, to name a few, demand integrated solutions that can make intelligent use of both cloud and edge systems, while applying machine learning and artificial intelligence tools to handle their growing complexity and to leverage the vast amount of available data created.
Sponsored by the IEEE Computer Society and IEEE Computer Society's TCDP, the 20th edition of the IEEE International Conference on Mobile Ad-Hoc and Smart Systems (MASS) was held in Toronto, Canada in September 2023, and brought together researchers, developers, and practitioners to address recent advances in mobile ad-hoc and smart systems, covering algorithms, theory, protocols, systems & applications, experimental evaluations and testbeds, security/privacy, as well as AI/ML-based smart design.
Congratulations on this achievment Yingying !
Third Annual Rutgers Robotics Workshop Successfully Held on September 22, 2023
Rutgers students won the first-place award of "Fair and Intelligent Embedded System Challenge" in Embedded System Week (ESWEEK 2023)
Guest Lecture - Prof. Basset, University of Pennsylvania
You're invited !
This Guest Lecture will be available via Zoom
Zoom link: https://rutgers.zoom.us/j/94481510460?pwd=VlVKNmw1Y0xuRkg4ODYra0FCNk9JUT09 (Passcode:21373)
WINLAB Chief Technologist Ivan Seskar in One of the Five Teams Selected by NSF $25 million Investment to Secure 5G Technologies
ECE students Spilios Evmorfos and Zhaoyi Xu win the Best Student Paper Award at IEEE MLSP
The ECE Department is proud to announce that the paper authored by Spilios Evmorfos and Zhaoyi Xu (under the supervision of Prof. Athina Petropulu) has received the Best Student Paper Award at the 2023 IEEE Workshop on Machine Learning for Signal Processing (IEEE MLSP), Rome Italy. The paper was presented at the MLSP workshop by Spilios. The abstract of the paper is given below.
Congratulations to Spilios, Zhaoyi, and Athina !
Spilios Evmorfos, Zhaoyi Xu, Athina Petropulu
The efficacy of sensor arrays improves with more elements, yet increased number of elements leads to higher computational demands, cost and power consumption. Sparse arrays offer a cost-effective solution by utilizing only a subset of available elements. Each subset has a different effect on the performance properties of the array. This paper presents an unsupervised learning approach for sensor selection based on a deep generative modeling.
The selection process is treated as a deterministic Markov Decision Process, where sensor subarrays arise as terminal states. The Generative Flow Network (GFlowNet) paradigm is employed to learn a distribution over actions based on the current state. Sampling from the aforementioned distribution ensures that the cumulative probability of reaching a terminal state is proportional to the sensing performance of the corresponding subset. The approach is applied for transmit beamforming where the performance of a subset is inversely proportional to the error between its corresponding beampattern and a desired beampattern.
The method can generate multiple high-performing subsets by being trained on a small percentage of the possible subsets (less than 0.0001% of the possible subsets for the conducted experiments).