Today, with many IoT devices and the development of Internet technologies, both wired and wireless connectivity play a crucial role in offering a wide range of services to end-users. Moreover, such communication and networking technologies have influenced individuals, institutions, governments, and industries. As a result, communication technologies have become an indispensable part of daily life. Recently, Machine Learning (ML) technologies have achieved breakthrough advances. When combined with communication technologies, the achievement in AI has created unique applications such as smart cities, self-driving cars, etc.
Our research group focuses on the study and development of high-speed, dependable, and flexible networking technologies, using theoretical and practical techniques to incorporate social, information, device, and energy factors. We study to leverage AI in enhancing and optimizing network performance. We’re also looking into solutions for analyzing and mining data collected by IoT devices. See the slides here for more detail.
Contact: Dr. Nguyen Phi Le, Email: firstname.lastname@example.org
Group website: http://icn.bkai.ai/
- Network optimization: Exploiting machine learning, deep learning, reinforcement learning, meta-heurisctic techniques to optimize network performance.
- Data mining in IoT networks: Focusing on processing, analyzing and mining data collected from IoT devices.
- AI in networking: Applying various machine learning and deep learning techniques including graph networks, deep reinforcement learning, etc., to analyze and predict the network behavior. Leveraging AI in controlling network operations.
- Distributed systems: Addressing different problems in distributed systems such as Synchronization, Replication and Consistency, Fault tolerance, etc.
- New generation network: Solving problems of NGN in expanding the existing access network infrastructure into networks capable of satisfying the user’s requirements.
- QoS-QoE: Proposing adaptive mechanisms to predict and optimize QoS/QoE in network systems.
- Internet of Things: Addressing different issues of IoT systems such as security problems, lack of regulation, limited bandwidth, etc.
- Optimization problems in Wireless Rechargeable Sensor Networks: We focus on a new wireless charging paradigm that considers a variety of factors such as charging path, charging time, target coverage, and connectivity. Furthermore, our system considers various optimization goals, including minimizing the number of mobile chargers, optimizing the depot placement, and extending the network lifetime. To solve the problem, we go from theoretical modeling to applying state-of-the-art technologies. Specifically, we leverage reinforcement learning and deep reinforcement learning, fuzzy logic, genetic algorithms to propose novel charging algorithms for wireless rechargeable IoT networks.
- Mobile air quality monitoring network: This research topic, part of the VinIF-funded Fi-Mi project (http://fi-mi.vn), presents a new mobile air quality monitoring network based on vehicle-mounted sensors. We explore various research issues, such as resource allocation, effective communication, data analysis, and data mining. The resource allocation problem is solved using optimization techniques such as reinforcement learning, fuzzy logic, and meta-heuristics. A mathematical model for opportunistic communication in mobile air quality monitoring networks is also proposed. We then leverage deep reinforcement learning for offloading data from devices to the servers. Finally, we use machine learning methodologies to calibrate the acquired data and propose deep learning models for spatiotemporal prediction problems.
- Satellite precipitation calibration: We investigate a fascinating research problem that is how to calibrate the satellite precipitation data. Our method uses deep learning techniques to blend satellite data that is fine-grained but inaccurate with the coarse-but-accurate data from ground-based monitoring stations. We then leverage graph neural networks to model the spatial correlation between the ground-based monitoring stations. Besides, recurrent neural networks are utilized to capture the data’s temporal properties.
- Prediction of water levels on rivers in Vietnam: This study focuses on data mining regarding river discharge and water levels in Vietnam. For hydrological research and flood prediction, forecasting river discharge and water levels have long been necessary. However, existing machine learning-based methods cope with severe flaws such as a lack of training data, noise in the obtained data, and the difficulty of altering the model’s hyper-parameters. In this research, we address at the same time these three challenges. Specifically, we apply data preprocessing techniques such as SSA to clean the data before feeding it into the forecasting models. Furthermore, we propose novel deep learning models that employ ensemble learning techniques and graph neural networks to capture both temporal and spatial correlation and extract relevant information from historical data. Furthermore, optimization techniques like metaheuristics are used to obtain the optimal hyper-parameters automatically.
- Traffic prediction and intelligent routing: In this research, we study how to learn and predict networks’ future behavior dynamically. From that, we design algorithms to control the network operations efficiently and intelligently. Several research problems are taking into account, such as exploiting the Graph Neural Networks to capture the spatial relation of the network and then leveraging deep learning to predict the network traffic; using reinforcement learning to guide the packets; utilizing metaheuristic algorithms and linear programming to optimize the routing paths. We perform experiments on network simulators such as OMNET, NS2, and real datasets such as Brain, Abilene, and Geant.
- Applying ML to Software-Defined Network => Knowledge-defined Network: Applying AI to the management of computer networks, making it to be a self-adaptive one. This means every network node will be an AI node, which is able to make decisions automatically. However, these network nodes have only a local view and a local control, hence, applying Artificial Intelligence to each network node is not trivial. For this purpose, a centralized control architecture, e.g. Software-defined Network (SDN), is a promising candidate. In this work, we decide to develop a knowledge layer on top of the SDN architecture, forming a Knowledge-defined Network.
- Replication and consistency in heterogeneous and Distributed SDN networks: Proposing an East-West interface, called SINA, to provide the interoperability of a heterogeneous and distributed SDN network. In addition, a novel Reinforcement Learning-based consistency algorithm is introduced for an adaptive Quorum-based replication mechanism.
- Applying Federated Learning to SDN: The learning phase of a multi-domain SDN network must address the following problems: both solutions of 1) learning separately in each domain and 2) collecting all data from different controllers and launching the training phase in a centralized server are not efficient due to the computational cost, training time, and the accuracy. That motivates us to apply Federated Learning, in which distributed SDN controllers cooperatively train a centralized learning model while do not share the training data for preserving data privacy.