This project has contributed a set of innovative approaches in exploiting machine learning (ML) techniques for wireless network optimization.  Some important insights have been revealed. 1) Machine learning algorithms, properly trained with optimization solutions from historical cases, can smartly identify critical features to expedite the computing of new instances. 2) To materialize such benefits, ML methods need to be integrated with proper domain knowledge on wireless network optimization to form innovative ML based optimization technologies. 3)  The disruptive techniques developed in this project can significantly mitigate the online computation overhead while maintain close-to-optimal performance, for the multi-hop wireless network optimization problem where the traditional optimization theories haven’t seen much progress in the recent decade. Specifically, the findings from this project have fundamentally advanced the state of the art of ML assisted network optimization from five aspects.

 1) We demonstrate that supervised machine learning, properly trained from historical optimization instances, can identify a subset of important links to form a smaller scale problem, which can generate close to optimal solutions with significantly less computation complexity. Furthermore, we incorporate graph embedding into deep learning (DL) to make our optimization techniques topology aware: the ML-assisted optimization framework trained over limited topology can generalize very well to new topologies.

2) We reveal that the widely adopted delayed column generation (DCG) optimization algorithm can be formulated into a reinforcement learning framework. Our inspiration comes from the insights that the iterated updating of independent sets (IS) for solution improvement in DCG is equivalent to a sequential Markov decision process (MDP) and thus enabling the reinforcement learning (RL) based optimization. The RL oriented vision then grants a disruptive method to further mitigate the computation overhead over DCG, which was not possible in the traditional optimization theory, by using ML to control the link weights for IS updating in each iteration. 

3) We develop an innovative self-supervised learning framework for accelerating wireless network optimization, where the trained machine can directly suggest an IS based scheduling structure (i.e., avoiding the iterated search as in traditional optimization decomposition algorithms) for a new optimization instance through an innovative similarity analysis. We fully implement the self-supervised learning framework that can realize a close-to-optimal approximation for the multi-commodity optimization problem with one round simple linear optimization.

4)    The design and materialization of our ML assisted optimization technology have stimulated the development of a series of machine learning implementation techniques, including but not limited to topology embedding, currirulum training, link prediction with attention, link differentiated loss function, scheduling structure pooling representation, dimension reduction, and scheduling structure classification.

5)  We demonstrate the efficiency of both supervised leaning and reinforcement learning in wireless network optimization through a few cases studies, including ML assisted capacity optimization for B5G/6G integrated access and backhaul Networks, DL assisted age of information optimizing over WiFi, and RL assisted energy optimization in multi-UAV based random access IoT networks.

This project seamlessly integrates studies in the areas of machine learning, optimization, stochastic control, graph theory, and wireless networking and advances the state of the art. Although the research community has applied ML to address many kinds of wireless networking problems, the study of ML for multi-hop wireless network optimization didn’t receive enough attention. It is expected that the technologies developed in this project can bring stimulating signals to the community that a ML-oriented perspective can indeed enable some disruptive methodologies to address some classic yet highly challenging network optimization problems.

This interdisciplinary research have provided many training projects to Illinois Institute of Technology (IIT) undergraduate and graduate students, students from Spain through IIT exchange program, and non-IIT college students and high-school students. Especially, this project inspired students (particularly PhD students) to pursue high-quality research with a creative, open-minded, and cross-disciplinary perspective. A couple of PhD students involved in this project had graduated and obtained R&D jobs in industry, and another on-going student has built up solid publication record with a promising career in academia.  

The outcomes from this project have been published or presented in top-tier IEEE conference proceedings and journals, including invited papers at the IEEE International Conference on Mobile Ad-Hoc and Smart Systems (MASS) and the Bronze Medal Award at the IEEE Communications Society (ComSoc) Excellence Camp co-located with IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC) 2023.  The publications have been timely listed in the project’s website. The outcomes have also been disseminated to community through community outreach activities and several invited/keynote talks at conferences, universities, or industry labs. The research results have also been leveraged to enhance the PI’s course materials at Illinois Institute of Technology. 

Last Modified: 01/09/2024
Modified by: Yu Cheng