Ongoing Projects

Network Resource Coordination with Explainable Learning Algorithms

Keywords: Network Slicing, Explainable AI, Online Convex Optimization

While network resources can be statically allocated to each network slice, it is further beneficial for both users and network operators to dynamically adjust the allocation and utilize the spare capacity efficiently. This type of network slice is called an elastic network slice. An indisputably important problem is the dynamic resource allocation under uncertain traffic trends. Furthermore, the profitability and scalability will be dominated by such allocation strategies when network functions on the slices collaboratively realize service functionality. This project aims to propose a resource allocation methodology to deal with the tradeoff between over- and under-provisioning for network slices hosting dependent network functions, guaranteeing the isolation of network slices.

ML-based Revenue Maximization for 5G and Beyond Networking

Keywords: Network Slicing, Mechanism Design, Neural Network-based Auction Design

An approach to deal with the uncertainty of the future demand of network resources is to shape the future by indirectly guiding the behaviors of network users through dynamic pricing. Common cloud computing platforms such as AWS have pricing mechanisms to solve similar problems for their computing resources. In this project, our goal is to propose a ML-based pricing mechanism that maintains an ideal distribution of slice request types in a communication network.

Asynchronous Federated Learning Platform

Keyword: Federated Learning

The advancement of ML techniques has broadened the potential use-cases of artificial intelligence. However, the transfer of huge amounts of data raises some critical issues including privacy-preservation and scalability. Federated Learning (FL), which transfers locally trained ML models instead of raw local data, is one of the promising directions of ML systems. Our proposal on a new FL platform demonstrates its uniqueness in avoiding the collection of all local ML models from FL users in a centralized location. The decentralized FL platform enhances the scalability of FL systems by partitioning FL users into multiple clusters that are each operated by a decentralized FL aggregator. The use of independently operating decentralized FL aggregators for disjoint FL user clusters also increases the edge computing-friendliness of the system. To realize the decentralized platform, the computing resources at the network edge need to be coordinated, considering latency and load-balancing.

 

 

Past Projects

Reliable Network Design considering Interdependencies

The increasing demand for diverse network services entails more flexible networks that are realized by virtualized network equipment and functions. Effective use of virtualization techniques requires sophisticated management of relations among softwarized modules since a set of the modules and hardware hosting them collaboratively provide a functionality. Several prior research works demonstrate that poor management of such inter-module relations could cause unique fragility of networks, called cascading failure. In a cascading failure, a failure of a part of a network propagates to the rest of the network along with the dependency relations among the modules and hardware. The focus of my work is to (i) identify dependency structures that can mitigate the impact of cascading failures and (ii) propose a protection method that constructs such robust dependency structures by appropriate resource allocation. The effectiveness of my protection method is demonstrated by theoretical results based on feedback vertex hitting and by experiments simulating different types of network failures.

G. Ishigaki, R. Gour and J. P. Jue, “Improving the Survivability of Clustered Interdependent Networks by Restructuring Dependencies,” in IEEE Transactions on Communications, vol. 67, no. 4, pp. 2837-2848, April 2019, doi: 10.1109/TCOMM.2018.2889983.

Deep Reinforcement Learning-based Recovery Resource Allocation for Dependent Network Functions

Recovery of failed dependent network components is a challenging task that involves additional complexity compared to the recovery of independent network components. In particular, the challenge becomes evident when the recovery is conducted progressively due to limited repair resources. The ordering of resource allocation in the recovery phase influences the interim computation and communication capability. This is because certain network components require not only repair resources but also the working operation of other network components in order to function properly. This combinatorial utility on resource allocation decisions makes the recovery problem more complex than typical network recovery problems. Hence, my project dealt with a progressive recovery problem under limited resources in networks with dependent virtualized network functions. I proved the NP-hardness of the problem and introduced a Deep Reinforcement Learning (Deep RL)-based technique that was able to achieve near-optimal solutions. Our simulation results also indicated that the proposed method is more robust to adversarial failures, compared to baseline algorithms. Furthermore, a set of our theorems show that a broad class of progressive recovery problems can be reduced to the cases that were solved in this project.

G. Ishigaki, S. Devic, R. Gour and J. P. Jue, “DeepPR: Progressive Recovery for Interdependent VNFs With Deep Reinforcement Learning,” in IEEE Journal on Selected Areas in Communications, vol. 38, no. 10, pp. 2386-2399, Oct. 2020, doi: 10.1109/JSAC.2020.3000402.