Introduction

The sixth generation (6G) of mobile communication systems is going to be massively connected and complex to fulfill the requirements of intelligent transportation and smart cities, industrial automation, and Augmented / Virtual Reality (AR/VR), among others. High-definition visual data are needed by these applications, demanding higher data rates and thus, the need for further enhanced Mobile BroadBand (FeMBB) services. High broadband data rates along with reliable and low latency communications (MBBRLLC) are essential for mission-critical tasks, such as autonomous driving and control of unmanned aerial vehicles; the reliable and low latency Machine Type Communication (RLLMTC) facilitates massive connectivity, sensing, and monitoring between a large number of devices. Conventional network management techniques that require human involvement are unfeasible due to the highly dynamic nature of these intelligent Internet-of-Things (iIoTs) services that require frequent sensing and monitoring, processing and computing of heterogeneous data types, and exchanging information or control decisions with large number of end devices and end subsystems. In this context, traffic prediction is crucial to perform these functions and satisfy application requirements. These applications also need significant computing power. Edge and fog computing architectures that extend cloud computing to the network edge and users terminals, respectively, reduce latency and increase reliability when performing computing tasks. Recent works in this area have identified the need to automate resource management and handle its complexity while reducing time decision making, but failed to provide up-to-date solutions to achieve reliable and real-time resource allocation. This project proposes to achieve this through continuous learning and adaptation, as a departure from traditional machine learning (ML) approaches as they are no longer suitable for high dynamic environments.

In fact, current ML use statically trained models which do not adapt to the traffic dynamics. These algorithms learn in isolation and even the most powerful solutions such as Deep Learning (DL) suffer from so-called catastrophic forgetting: they forget previously-learned information upon learning new one. Learning every task independently (e.g., finding available spectrum in different frequency bands, available access points in different operator networks) is not scalable, and it requires a training phase after which no further learning occurs. Thus, DL is not able to provide resource allocation “on the fly”, even when resources are in fact available. Continual and reinforcement learning algorithms are instead needed to learn continually or lifelong on the job during model application to adapt to the changing trends. Lifelong ML (LML) is an advanced machine learning paradigm that does not stop learning from its experience, and is not bound to a training phase. LML algorithms present human-like intelligence and are able to learn continuously, accumulate knowledge and adapt to learn in open environments. Although some progress has been made to make learning models respond faster to dynamic environments by transfer learning and lifelong learning, they have shown limitations when applied to learning different tasks. Besides, current implementations are mostly centralized, increasing latency, and overhead traffic in the network. In addition, only a limited amount of informative data is available at the network edge/fog (generated by applications, sensing devices, users, network, etc.) to perform accurate traffic prediction, and exchanging raw data between multiple devices is constrained to their battery and availability of network resources. Therefore, the complexity and dynamics of wireless networks require new distributed LML algorithms. When modeling data are exchanged, automated knowledge transfer across network edges under privacy concerns is required. LML is able to achieve that by transferring knowledge through models rather than data. Due to the hierarchical design of wireless networks, different degrees of data accessibility, privacy concerns, and availability of resources (communication, computation, and storage limitations on the network as well as on-devices) exist at different parts of the system. Artificial Intelligence (AI) must adapt its mode of operation from centralized to distributed, static to dynamic, personalized to generalized, to be energy-efficient and robust against the spatial and temporal dynamics in the network.

Research Thrusts

This project aims to address these challenges and advance the design of intelligent, reliable, and self-reconfigurable networks capable of real-time resource management for massive connectivity in sensing, communication, and computation within iIoT networks. To cope with the complexity and heterogeneity in wireless networks, we develop a cross-system architecture design methodology that enables the design of future dynamic wireless networks by optimal integration of information, communication and AI. Our collaborative research will be conducted through the following thrusts.

Thrust 1: Cross-system Architecture Design

We design CRUISE, a cross-system architecture that enables optimal integration of Information (I), Communication (C) and Artificial Intelligence (AI) Technologies (ICAIT). The major challenge to jointly optimize ICAIT is to develop a tractable framework that captures their interdependency and tradeoffs for different use cases (e.g., FeMBB, MBBRLLC, RLLMTC). We will minimize the regret (system losses) due to delay and imperfections in the information, as a compromise between complexity and accuracy of the predictions. We also investigate how to model centralized learning based on LML when edge devices can only use limited processing power and the architectural support needed for Lifelong Learning-enabled Wireless Networks (L2WN). For this reason, we will build on and extend the standard reference architecture released recently by the Open Radio Access Network (O-RAN) Alliance [63].

Thrust 2: Autonomous Reconfiguration in L2WN

We develop ICAIT adaptation schemes to autonomously reconfigure the connections due to dynamic availability and demands of network resources at the wireless edge. The goal is to learn which nodes are the most relevant to identify the areas with the highest expected traffic and allocate most of the resources to guarantee connectivity. Then we take another step further in the development of LML and investigate a distributed LML for mission-critical decision making under information and communication dynamics. We require the LML to perform transfer learning to efficiently learn tasks in a new environment (e.g., available resources in different frequency bands, available access points in different operator networks) and we assess its accuracy.

Thrust 3: Robustness and Energy Efficiency

The robustness of the proposed ICAIT learning framework is analyzed in terms of reliability along two different dimensions: communication and reliability of inference under low-power limitations. We develop connectivity probing and reinforcement schemes to check the availability of the connections and reinforce (strengthen) them to meet stringent reliability and latency requirements under traffic dynamics. We also study traffic forecasting mechanisms to reduce the energy consumption by switching base stations to a low-power state in periods of low utilization using the LML algorithms developed in the previous thrusts. Our solutions will be validated by a prototype system implementation and real-world experimentation using the COSMOS platform (USA) and the 5G Test Network (Finland).

Educational Activities and Outreach

The outcomes of the research are integrated in the following courses:

• ECE 597/697 AI-based Wireless Networks Design (University of Massachusetts Amherst)
• ECE 325 Computer Networking (University of Massachusetts Amherst)
• CS538 Data Mining and Text Mining (University of Illinois Chicago)
• CS411 Artificial Intelligence (University of Illinois Chicago)
• CS-E4100 Seminar in Computer Science: Internet, Data and Things (Aalto University)

Additional outreach aims to increase the general awareness of engineering issues in the K-12 community. The huge number of smart home IoT devices makes our project ideal for introducing the basic concepts of wireless communication and the need of AI, especially to high-school students and junior college students, including through the Aalto Junior initiative.

There is a plan to organize a virtual summer school to disseminate the results of the projects with industry participation from Finland (Oulu’s 6G flagship) and US (Juniper networks and MITRE Corporation, among others).