NeTS: Small: Machine Learning Meets Wireless Network Optimization: Exploring the Latent Knowledge, 09/01/18-08/31/22

Major Goals of this Project

With the superior capability of discovering intricate structure of large data sets, machine learning and one of its branch, deep learning, have been widely applied in various areas including wireless networking. While existing deep learning applications mainly focus on data analysis and classification tasks, the role it can play on fundamental research issues in wireless networks, such as optimization and resource allocation, is yet to be explored. With the proliferation of heterogeneous wireless networking infrastructure and mobile applications, wireless network optimization has seen a tremendous increase in problem size and complexity, calling for a paradigm for efficient solution. This project aims at a pioneering study on how to exploit deep learning for significant performance gain in wireless network optimization, through three complementary research tracks: link evaluation through learning, critical transmission pattern identification and online scheduling facilitated with learning. This project will contribute to answer the following fundamental questions: what knowledge can be extracted from computing experiences of network optimization problems? How to obtain this information? How to exploit the information to benefit wireless network optimization?

Intellectual Merit

The proposed research is expected to contribute innovative approaches in exploiting machine learning techniques for wireless network optimization, with fundamental intellectual merits in four aspects. 1) This study will reveal the insight that machine learning algorithms, properly trained with optimization solutions from historical cases, can smartly identify critical features (a small set of critical links or transmission patterns) that lead to optimal or close-to-optimal solutions for a new case. 2) Identifying critical features (i.e., the latent knowledge) that can facilitate wireless network optimization is different from the classification operation in common machine learning application scenarios; this project will construct the learning framework and systematically design related functions or algorithms that can accommodate the unique needs and requirements of the optimization tasks in wireless networks. 3) This study will demonstrate how the learned knowledge can be exploited to significantly reduce computational complexity in wireless network optimization and online scheduling. 4) With low-complexity optimization enabled by machine learning, it is then possible and convenient to analyze the dynamic relationship between the input data traffic and internal network features (link or pattern activation) or resource allocation (scheduling or routing); such findings and insights may have deep impact on network resource allocation methods.

Research Achievements

1) Research Activities in Year 1

• Topology-aware machine learning for wireless network optimization --- we propose a topology-aware learning based framework for wireless optimization problems by incorporating graph embedding techniques with supervised deep learning.

• Effective machine learning implementation techniques in wireless network optimization --- Through the training and testing studies in wireless networks based on the protocol interference model, we demonstrate that machine learning efficiency can be improved through "weight differentiated cost," "optimal batch size for training," and "optimal threshold for link selection."

• Deep reinforcement learning for D2D link scheduling --- In this work, we study a representative network optimization task of maximizing the throughput-based system utility through link scheduling in a single-radio, single-channel D2D networks, and propose a novel learning-based framework to extract the efficient scheduling policy.

2) Research Activities in Year 2

• Deep reinforcement learning for for scheduling in multi-hop wireless networks --- In particular, we consider the canonical multi-commodity (MCF) problem in multi-hop wireless networks and leverage the methods of deep reinforcement learning to generate scheduling policies, addressing the challenges of the involvement of non-differentiable operations and the consideration of variable network topologies. .

• Machine learning for energy sustainable multi-UAV based random access IoT networks --- In this work, we apply the Non-Orthogonal Multiple Access (NOMA) technique to improve the massive channel access of a wireless IoT network where solar-powered Unmanned Aerial Vehicles (UAVs) relay data from IoT devices to remote servers. Specifically, IoT devices contend for accessing the shared wireless channel using an adaptive p-persistent slotted Aloha protocol; and the solar-powered UAVs adopt Successive Interference Cancellation (SIC) to decode multiple received data from IoT devices to improve access efficiency. In such a system, we leverage machine learning to tackle the joint problem of dynamic multi-UAV altitude control and multi-cell wireless channel access with multiple energy constraints.

• Enhancement to topology-aware machine learning techniques --- In this work, we have developed some important machine learning implementation techniques that can further enhance our topology-aware deep learning framework (TADL) to achieve a robust performance in predicting important links over various network topologies with different sizes, including currirulum training, link prediction with attention, and link differentiation loss function.

3) Research Activities in Year 3

• A self-supervised learning approach for accelerating wireless network optimization --- One basic idea of this research is extracting knowledge from historical problem instances to accelerate the solution of new instances. Although this idea is intuitive, its implementation is by no means trivial. To achieve this goal, two fundamental challenges need to be addressed. The first one is how to effectively identify the similarity between two problem instances. The other challenge is the proper usage of the historical instances. Even with a means to identify the similarity between problem instances, directly applying the old instances' solutions is unlikely to give good performance, neither is it practical to store all the past solutions due to the huge size of input configuration space. We have developed an innovative self-supervised learning based approach to address these two challenges.

• Deep reinforcement learning for enhancing the delayed column generation method --- The delayed column generation (DCG) method is well-known as an effective approximation algorithm to solve the multi-hop wireless network optimization problem, although it may incur a large number of iterations in certain cases. This year, our study achieved an insightful understanding why DRL can further enhance the DCG performance. With an innovative vision of viewing column search as a routing issue in the multi-dimensional solution space of the optimization problem, we figure out the essence is that the traditional DCG is always searching paths in a greedy manner at each iteration, while the DRL focuses on global optimality and may take sub-optimal actions at certain iterations. Such insights guide us to improve our design in a systematic manner and enhance the performance.

• Understanding the supremacy of deep learning in wireless networking --- In this work, we systematically reviews recent attempts at leveraging DL for addressing wireless network optimization problems, presenting a fundamental understanding of where and how the supremacy of DL comes regarding the wireless network optimization. Our vision is that DL technologies can advance the state of the art of wireless network optimization from four aspects: establishing informative formulations of the optimization tasks; alleviating the computational overhead of complex problems; enabling disruptive algorithm design; and discovering new latent knowledge facilitating optimization.

4) Research Activities in Year 4

• A topology-aware approach for exploring and exploiting the scheduling structure --- While the self-supervised learning framework was started from the study of a given fixed topology, we have successfuly enhanced this powerful framework with a topology-aware capability: the trained model can be used for optimization over different network topologies. The foundational challenge comes from the representation for both inputs and outputs of the model: different topology may incur different dimensionality. We develop an innovative geometric canonical representation (GCR) system to address these challenges.

• Effective scheduling exploitation in high-dimensional space --- In the self-supervised learning framework, the SSC stage needs to be conducted in a high-dimensional vector space. Each class contains a subset of independent sets (ISs) selected from the exponentially large space. Furthermore, a certain IS is very possibly involved in multiple classes. The generic topology-aware settings makes the solution space larger in orders. An exhausted search of all possible scheduling classes is then out of the question. We develop effective dimension reduction and novel cross-class exploitation techniques to tackle these issues.

• A deep learning assisted approach for optimizing information freshness in a WiFi network --- We leverage ML to optimize information freshness for remote monitoring over a practical WiFi network, where a tagged node under study needs to deliver status messages to the monitoring application installed at the access point (AP).We model the tagged node as a first-come-first-serve (FCFS) queue over the MAC layer and optimize the information freshness (equivalently, the age of information) by adjusting the traffic arrival rate (i.e., the sampling rate of the monitoring node) to meet an optimal utilization factor. We develop a deep learning (DL) based approach to predict two critical probabilities which are required for accurate MAC service time analysis. A particular original design is that our DL module only needs to use the data that is readily available at the AP and the tagged node, incurring no communication overhead with other nodes in the network.

5) Research Activities in Year 5

• Proper super-parameters to improve the performance of self-supervised learning framework --- The self-supervised learning framework involves many super-parameters impacting performance. We study the setting of super-parameters from three aspects. 1) Instance representation. We define a "pooling" operation to summerize all the independents sets associated with the optimal solutions of an instance into a single vector representation. 2) Dimension reduction. We find that the principle component analysis (PCA) based dimension reduction perform very well for our problem while significantly mitigate the complexity. 3) Clustering into scheduling classes. We compare the K-means++ algorithm with the standard K-mean and find the former one leading to better performance. We also conduct extensive experiments to determine the proper number of class (i.e., the value of K) that robustly perform well in our setting.

• Self-renewal machine learning approach for improved learning performance --- While our self-supervised learning framework has reasonable generalization capability, but it is up to a capacity bound. If in practice the number of commodity flows is much more than that used in the training data, the optimization performance based on the scheduling structure suggested by the AID will degrade. We envision that the cases for which the AID machine in our framework has the worst performance in identifying the scheduling structure will be the most effective training data instead for performance enhancement. Such an idea was materialized with a simple yet effective threshold based self-renewal training method. Extensive numerical experiments demonstrate the efficiency of this method in performance enhancement compared to the standard retraining with non-selective samples.

• Machine learning assisted capacity optimization for B5G/6G integrated access and backhaul Networks --- The millimeter-wave (mmWave) gigahertz (GHz) spectrum has been considered in beyond 5G (B5G)/6G networks to provide extra capacity. However, mmWave connection is susceptible to interference and degradation, causing coverage limitation, and thus requires the denser deployment of multi-hop wireless backhauling to connect densely depolyed small base stations (BSs). The wireless backhaul solution proposed by the third generation partnership project (3GPP) Release 16 is the integrated access and backhaul (IAB) architecture. In such a B5G/6G IAB network, the multi commodity flow deployment problem between the small BSs and the macro-BSs is very suitable for the ML assisted optimization framework, which has been developed through this project. We thus consider this part as a case study to demonstrate the efficacy of our method.

Publications

• O.T. Ajayi, X. Cao, H. Shan, and Y. Cheng, "Self-renewal machine learning approach for fast wireless network optimization," in Proc. The 20th IEEE International Conference on Mobile Ad-Hoc and Smart Systems (MASS 2022), September 25 - 27, 2023, Toronto, Canada.

• O.T. Ajayi, S. Zhang, and Y. Cheng, "Machine learning assisted capacity optimization for B5G/6G integrated access and backhaul networks," in Proc. IEEE INFOCOM 2023 Workshop on Pervasive Network Intelligence for 6G Networks (PerAI-6G), virtual, May 2023.

• S. Zhang, O.T. Ajayi, and Y. Cheng, "A self-supervised learning approach for accelerating wireless network optimization," IEEE Transactions on Vehicular Technology, vol. 72, no. 6, June 2023, pp. 8074-8087.

• S. Zhang, B. Yin, W. Zhang, and Y. Cheng, "Topology aware deep learning for wireless network optimization," IEEE Transactions on Wireless Communications, vol. 21, no. 11, pp. 9791-9805, November 2022.

• S. Wang and Y. Cheng, "A deep learning assisted appraoch for minimizing the age of information in a WiFi network," in Proc. IEEE MASS 2022,invited paper, October 20 - 22, 2022, Denver, Colorado

• Z. Cheng, R. Zhang, L.X. Cai, Y. Cheng and Y. Liu, "A deep reinforcement learning based approach for NOMA-based random access network with truncated channel inversion power control," in Proc. IEEE ICC 2022,16-20 May 2022, Seoul, South Korea (Hybrid)

• J.M. Parella, O.T. Ajayi, and Y. Cheng, "Adaptive messaging based on the age of information in VANETs," in Proc. IEEE GLOBECOM 2022, 4-8 December 2022, Rio de Janeiro, Brazil (Hybrid)

• S. Wang, Y. Cheng, L.X. Cai, and X. Cao, "Minimizing the age of information for monitoring over a WiFi network," in Proc. IEEE GLOBECOM 2022, 4-8 December 2022, Rio de Janeiro, Brazil (Hybrid)

• J. Wang, X. Cao, B. Yin, and Y. Cheng, "Sleep-Wake Sensor Scheduling for Minimizing AoI-Penalty in Industrial Internet of Things," IEEE Internet of Things, vol. 9, no. 9, pp. 6404-6417, May 2022.

• S. Zhang, B. Yin, and Y. Cheng, "Deep reinforcement learning for scheduling in multi-hop wireless networks," invited paper, in Proc. The 18th IEEE International Conference on Mobile Ad-Hoc and Smart Systems (MASS 2021), Virtual Conference, October 4-7, 2021.

• Y. Cheng, B. Yin, and S. Zhang, "Deep learning for wireless networking: The next frontier," IEEE Wireless Communications,vol. 28, no. 6, pp. 176-183, December 2021.

• S. Khairy, P. Balaprakash, L. X. Cai, and Y. Cheng, "Constrained deep reinforcement learning for energy sustainable multi-UAV based random access IoT networks with NOMA," IEEE Journal on Selected Areas in Communications, vol. 49, No. 4, pp. 1101-1115, April 2021

• S. Zhang, B. Yin, Y. Cheng, L. X. Cai, and S. Zhou, Z. Niu, and H. Shan, "Energy-Efficient Massive MIMO With Decentralized Precoder Design," IEEE Transactions on Vehicular Technology, vol. 69, no. 12, pp. 15370-15384, December 2020.

• B. Yin, S. Zhang, and Y. Cheng, "Application-oriented scheduling for optimizing the age of correlated information: A Deep Reinforcement Learning based Approach," IEEE Internet of Things Journal, vol. 7, no. 9, pp. 8748-8759, September 2020.

• M. Han, S. Khairy, L. X. Cai, Y. Cheng, and R. Zhang, "Reinforcement learning for efficient and fair coexistence between LTE-LAA and Wi-Fi," IEEE Transactions on Vehicular Technology, vol. 69, no. 8, pp. 8764-8776, August 2020.

• M. Han, S. Khairy, L. X. Cai, Y. Cheng, and F. Hou, "Capacity analysis of opportunistic channel bonding over multi-channel WLANs under unsaturated traffic," IEEE Transactions on Communications, vol. 68, no. 3, pp. 1552-1566, Mar. 2020.

• S. Zhang, B. Yin, S. Wang, and Y. Cheng, "Robust deep learning for wireless network optimization," in Proc. IEEE ICC 2020, June 7-11, 2020

• S. Zhang, W. Shen, M. Zhang, X. Cao, Y. Cheng, "Experience-driven wireless D2D network link scheduling: A deep learning approach," in Proc. IEEE ICC 2019, Shanghai, China, May 20-24.

• B. Yin, S. Zhang, Y. Cheng, L. X. Cai, Z. Jiang, S. Zhou, and Z. Niu, "Only those requested count: proactive scheduling policies for minimizing effective age-of-information," in Proc. IEEE INFOCOM 2019, Paris, France, April 29 - May 2.

Broader Impact Activities

• Three invited papers were published in the 18th, 19th, and 20th IEEE International Conference on Mobile Ad-Hoc and Smart Systems (MASS 2021-2023), respectively. IEEE MASS is a premier annual forum for sharing original, novel ideas in mobile ad-hoc networks and smart systems. Invited papers in such a prestigeous venue demonstrate the impact of our research.

• The PI gave a keynote talk titled as "To Find the Supremacy of Machine Learning in Wireless Network Optimization" at the Workshop of Edge Intelligence for 6G Networks, co-located with the IEEE/CIC International Conference on Communications in China (ICCC), Dalian, China, August 10-12, 2023. The talk contains some outcomes from this project. On August 14, the PI gave the same talk as an invited seminar speaker at the Department of Computer Science and Information Engineering, National Taiwan University. On September 26, the PI gave this invited talk again at University of Waterloo, Canada, as an IEEE Vehicular Technology Chapter presentation. The PI gave an invited talk "Machine Learning Meets Wireless Network Optimization: Exploring the Latent Knowledge" in IEEE International Conference on Computing, Networking and Communications (ICNC 2020).

• The PI has attended the IEEE Communications Society (ComSoc) Excellence Camp co-located with IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 5 September 2023, Toronto, Canada. The IEEE ComSoc Excellence Camp program (https://www.comsoc.org/education-training/ieee-comsoc-excellence-camps) was recently created to provide students, postgraduate students and young professional members with opportunities to stay up to date on the emerging technologies. IEEE ComSoc Excellence Camps aim to offer an intensive and immersive educational opportunity through lectures, panels, projects, and networking sessions in a succinct one to two day event. The PI's student Suyang Wang's project "A Deep Learning Assisted Approach for Minimizing the Age of Information in a WiFi Network" was presented in this ComSoc educational camp and won the Bronze Medal Award.

• In 2022, the PI's PhD student Oluwaseun Ajayi won the Illinois Tech Socially Responsible Modeling, Computation, and Design (SoReMo) Scholarship award. SoReMo provides unique interdisciplinary research and outreach opportunities for students who want to apply computational, modeling, and design skills to solve a social issue. Oluwaseun's SoReMo award is largely credited to his solid background on machine learning built up through his research activities in the NSF projects.

• A tutorial article "Deep learning for wireless networking: The next frontier" was published in IEEE Wireless Communications, 2021.

• The PI leveraged the studies in his NSF projects to enhance the research components of his courses, "ECE541: Computer Network Performance Analysis" and "ECE545: Advanced Computer Networks." Two week of lectures about applications of probability and queueing theory in AoI analysis, medium access control analysis, and machine Learning were added into the ECE541 course contents. ECE545 course contents were enhanced with discussions of AoI related studies in next generation wireless networks. The PI has been developing a new course "Machine learning in Wireless Networking," which is expected to deliver in 2024 for both graduate students and senior undergraduate students. In Fall 2023, one student, Pin-Chien Chen, was attracted to do his Master thesis research, with a focus on machine learning based techniques to detect fake news. Another student, Yamen Ziadeh, did a ECE597 special project on mobile edge computing under my supervision.

• The PI offered research projects to both undergraduate students and high school students. Since Summer 2022, a grade-11 student, Hannah Ding from Walter Payton College Preparatory High School, joined the PI's lab to do volunteer research. Hannah was teamed with Oluwaseun to conduct ML related research. A 3rd-year undergraduate student, Raymond Zhao, from the Computer Science Department, University of Wisconsin-Madison, started volunteer research in the PI's lab in Fall 2023 and currently conducts literature survey on ML in networking and ML security. In 2022, the PI also served as an external mentor to another grade-11 student, Matthew Wojcik from Bartlett High School, for his school research project. In summer 2019, the PI offered a research project on device-to-device networking to a minority undergraduate student Chimemezue Clinton (from ECE department, IIT), and offered a research project on machine learning to Sriram Ashokkumar, a junior student at Metea Valley High School, Aurora, Illinois.

• In Summer 2021, the PI worked with IIT Summer Engineering Undergraduate Research Immersion Program and offered a research course and offered a research course "ENGR498: Machine Learning in Wireless Network Optimization." The course was advertised by IIT and open to both IIT undergraduate students and visiting students from other universities or other countries. The course was successfully delivered. In the course, students used open-source machine learning tools to solve a scheduling problem in a device-to-device wireless networking context. In Fall 2022, he contributed "Optimizing Information Freshness in Internet of Things (IoT) Systems" to the Illinois Tech Armour R&D Research Project Program, which is offered to both undergraduate and graduate students.

• The PI conducted effective outreach activities through collaborating with IIT exchange program. . In Fall 2021 and Spring 2022, the PI supervised one visiting master student (Jordi Marias I Parella) from Polytechnic University of Catalonia, Spain. The student was teamed up with another PhD student (Oluwaseun Ajayi) and fruitfully conducted the research on AoI controlling in VANETs. One exchange student (home univerity being the Polytechnic University of Madrid, Spain) from the Department of Mechanical Engineering (different from the PI's department) was attracted to do master thesis research under the PI's supervision in Summer and Fall 2020. Another visiting student is doing research in the PI's lab in Fall 2021. Such outreach collaboration demonstrates the visibility of the PI's research group.

• The department of Electrical and Computer Engineering, Illinois Institute of Technology annually held a research event, ECE Day, yearly. ECE students from all degree levels exhibited their research work with posters. ECE external advisory board members (from industry and other universities) and some IIT alumni also attended the ECE Day event. Research students supervised by the PI presented their research supported by NSF funds. In 2019, 2022 and 2023, Shuai Zhang, Oluwaseun Ajayi and Suyang Wang, co-supported by this project, won the 1st, 3rd, and 2nd places of the PhD poster award in the area of computer networks and communications, respectively. Through this event, the PI's NSF projects and related research contributions had been well disseminated to the community.

• In Fall 2022, the PhD students in this project attended the IIT Welcome Week Student Research Poster Event and displayed the research contributions from the PI's NSF projects with research posters. This is a big event with university-wide attendance of undergraduate/graduate students and faculty members, alumni and family members, and invited judges from industry.

• In 2019 Summer, the PI organized a one-day mini-workshop on machine learning. All research students in the PI's lab participated and gave presentations. The presentations from students gave a good introduction to major types of modern machine learning technologies and studies several important academic papers regarding applications of machine learning in wireless networks, network security, and block chain.

Personnel Involved in This Project

• PI: Yu Cheng

• RA: Shuai Zhang

• RA: Md Tahmid Yasar

• RA: Bo Yin

• RA: Suyang Wang

• RA: Oluwaseun Ajayi

Computation and Simulation Codes

Available based on request.

---

[ Home | Research Projects | Publications | Honors & Awards | Teaching | Students | Professional Activities | Invited Talk ] 

---

FUTURE NETWORKING RESEARCH LABORATORY                      (under development)

Last Updated: October, 2021