Continuous Federated Learning of Global Policies in Coalition Environments

Abstract In coalition environments, multiple parties may need to collaboratively solve a coalition-based (or global) decision- making task in a federated manner, by using models trained autonomously from their own local datasets, and sharing as little information as possible. Instead of adopting a centralised federated learning approach, which may suffer from a single central server bottleneck, we propose a global policy learner framework that learns global policies by taking into account just the predicted outcomes of the locally trained models, hence requiring minimal information sharing among the local parties. A global decision-making task on new unlabelled data is solved by (i) collecting local predictions from the parties, through their own learned models, together with associated level of confidence, (ii) combining them to produce a label for the data, and (iii) using the symbolic learner FastLAS to learn from this labelled data a global policy capable of performing global predictions. Such policies can be continuously updated as new data become available and/or local models get retrained when needed. The learned global policies can be used in situations where the communication with the parties is unreliable due to failures or out-of-range drop out.
  • Yaniv Aspis (Imperial)
  • Daniel Cunnington (IBM UK)
  • Mark Herbster (UCL)
  • Alessandra Russo (Imperial)
  • Krysia Broda (Imperial)
  • Jorge Lobo
  • Ankush Singla (Purdue)
  • Elisa Bertino (Purdue)
  • Dinesh Verma (IBM US)
Date Sep-2020
Venue 4th Annual Fall Meeting of the DAIS ITA, 2020