Distributed Learning and Cooperative Optimization for Multi-agent Autonomous Wireless Access Points (DISCO)
DISCO receives funding from the Research Council of Norway through the program IKTPLUSS. The project is owned by SimulaMet and led by Chief Research Scientist and Research Professor Baltasar Beferull-Lozano, Head of the Department of Signal and Information Processing for Intelligent Systems (SIGIPRO) at SimulaMet.
The partners in the project are the University of Oslo (UiO), Altibox, Multinett, Bell Labs, Royal Institute of Technology (KTH) and Pompeu Fabra University (UPF).
The overall objective of DISCO is to design a new generation of decentralized data-driven AI-based methods for distributed learning and cooperative optimization over autonomous intelligent network ecosystems, with a special focus on large-scale autonomous cooperative wireless Access Points (APs).
Even though Wi-Fi networks are the primary medium for global internet traffic, their operation is far from optimal because of the lack of coordination between the different interfering Wi-Fi Networks. They have evolved tremendously, bringing new features not previously present, however, they come together with a very large plethora of configuration parameters, which makes them extremely complex to optimize dynamically.
Most of solutions proposed for network resource allocation are based on models and heuristics, not data-driven, and can not find efficiently the optimal configuration of all the parameters jointly across all the APs, due to the enormous search space. The few existing ML-based algorithms assume a centralized controller, which is not possible in practice for managing separately owned Wi-Fi networks, implying also a privacy threat.
DISCO novelties:
1) Novel distributed online ML algorithms to learn a multi-layer and multi-modal time-varying graph representing the overall dynamics of the interfering Wi-Fi networks, by using multivariate data time-series from APs & clients, detecting clusters of interfering Wi-Fi networks, allowing also to detect anomalies and predict variables.
2) A fully distributed Graph-Time Neural Network-based Multi-Agent Reinforcement Learning (MARL) algorithm for resource allocation, which operates directly on the learned graph.
3) Optimization of interplay between the protocol connecting the APs (while enforcing privacy), and the efficiency of the distributed MARL algorithm. The solutions will be implemented using portable containerized software modules and demonstrated in collaboration with key leading national and international stakeholders (Altibox, Multinett, Bell Labs).
The theoretical prescriptions and algorithms designs will also provide insight into designing similar distributed methods in other environments involving multiple agents that interact with each other and with the environment, to learn collectively and perform cooperatively complex tasks of interest in other application domains, such as distributed intelligence for teams of robots and wind farms.