Beyond 5G Network Slicing Based on Incremental Learning
Abstract
Tactical networks deployed to support military missions, disaster management, and/or emergency situations will benefit greatly from 5G technology and its efficient protocols for resource allocation. To achieve this, network slicing (NS), through the creation of functional chains, will be a key component in properly allocating resources and guaranteeing a desirable bandwidth and reliability. Most advances in NS have been in structured networks, where the assets are fixed or at least have a constant power supply. In sharp contrast, the performance of NS for unstructured networks, which are built on top of moving infrastructure capabilities (helicopters, tanks, unmanned vehicles) with Mobile Edge Computing (MEC) can still be further improved. The problem to be solved is twofold. On the one hand, unstructured networks are harder to predict, so there is a need for more accurate prediction models and less room for heuristics. On the other hand, since the network can easily become fragmented, the "control unit" of the resource allocation itself needs to be properly decentralized to avoid losing access to the allocation mechanism altogether. This proposal puts forward the use of advanced machine learning (ML) techniques: both supervised and unsupervised, to improve the allocation of virtual slices in tactical networks which shall be referred to as beyond 5G (B5G). This work has been organized into three logical sets. First, the design and development of advanced machine learning algorithms to solve the optimal allocation problem in possibly fragmented mobile networks. Both supervised and unsupervised methods will be tested for this purpose. For former methods, annotated data from past usage of resources: memory, CPU, and storage will be crucial, but also the conditions of the MEC devices spanning the network: location, power consumption, etc. For the latter, expertise in Markov Decision Processes and stochastic systems will allow the use of model-based reinforcement learning approaches, if data is not widely available. Secondly, since the network is formed between moving units, its nature is mostly dynamic. As a consequence, any machine learning model deployed needs to take into account this variability. For example, the topology of the network, often assumed known and fixed, will change between the snapshot used for training and the actual network seen during inference. In the ML community, this is known as concept drift: the distribution of the data changing after training. In this regard, incremental machine learning methods will be developed to guarantee an optimal allocation of resources regardless of the state of the network, as long as it is technically feasible to do so. To achieve this, constant and accurate monitoring of the network will be paramount. Finally, this project will explore the benefits of adding explainability to the ML system described above. It is expected that understanding how the system works will not only be beneficial for post-mortem analysis but could also influence strategy. It should allow the officers interfacing with the system to reason about counterfactuals in their strategy and how it will impact the quality of the network. The field of explainable AI (eAI) is still in constant development and this proposal will require the use of state-of-the-art methods applied to both supervised and unsupervised ML.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 08, 2023
- Source ID
- W911NF2310088
Entities
People
- Leandros Tassiulas
Organizations
- Army Contracting Command
- United States Army
- Yale University