Self-Adaptive Distributed Systems Paradigms:

The vision of this project requires new approaches to building distributed systems, in which self-adaptive behaviours are at the core of the system’s capability. Individual pieces of code will be highly mobile, able to move from the cloud to the edge and back, or even be resident on programmable SDN network devices where needed, depending on which permutation brings the highest benefit at any given time. This will demand new approaches to system construction, state management, and overall system management paradigms – all of which must provide guarantees of soundness for production-class systems.

Distributed Reinforcement Learning:

Our overall platform will represent a very large distributed search space for real-time reinforcement learning, in which one node must work with all other nodes to take local actions that contribute to a global reward. This represents a point in the real-time learning field which has received very little research for real-world systems. This track of research will investigate the state of the art in distributed reinforcement learning and seek to develop new techniques which scale well for state-of-the-art real-world stream processing systems.

Orchestration Methods, Models, and Languages:

A highly-adaptive distributed system covering thousands or tens of thousands of devices requires novel methods of orchestration so that system administrators can set policies and guide autonomous deployment decisions in desirable directions from a business perspective. This research track will therefore examine how orchestration models or languages which can be understood by administrators or even business executes, and which can be used with a highly adaptive deployment infrastructure to provide policies, guidelines, and constraints while still supporting high degrees of autonomous freedom.

Nation-Scale Analytics and Monitoring:

All of the above tracks are based on the ability to analyse and monitor the current state of a very large scale system, and to feed observations into orchestration and learning. This track will develop novel methods around how the network and distributed system is analysed on a continuous basis, at very large scale, and in a way that presents the information that is really needed by learning and orchestration. This includes adaptive data aggregation techniques to filter unneeded data, adaptive classifiers which can extract the right information to inform learning decisions, and human presentation techniques through which human users can explore the data and potentially suggest new classification techniques or trend analyses.