Authors:
Ladislav Jirsa
1
;
Lenka Kuklišová Pavelková
1
and
Anthony Quinn
2
Affiliations:
1
Institute of Information Theory and Automation, The Czech Academy of Sciences, Prague and Czech Republic
;
2
Institute of Information Theory and Automation, The Czech Academy of Sciences, Prague, Czech Republic, Trinity College Dublin, The University of Dublin and Ireland
Keyword(s):
Fully Probabilistic Design, Bayesian Filtering, Uniform Noise, Knowledge Transfer, Predictor, Orthotopic Bounds.
Related
Ontology
Subjects/Areas/Topics:
Adaptive Signal Processing and Control
;
Informatics in Control, Automation and Robotics
;
Sensors Fusion
;
Signal Processing, Sensors, Systems Modeling and Control
Abstract:
The paper presents an optimal Bayesian transfer learning technique applied to a pair of linear state-space processes driven by uniform state and observation noise processes. Contrary to conventional geometric approaches to boundedness in filtering problems, a fully Bayesian solution is adopted. This provides an approximate uniform filtering distribution and associated data predictor by processing the involved bounds via a local uniform approximation. This Bayesian handling of boundedness provides the opportunity to achieve optimal Bayesian knowledge transfer between bounded-error filtering nodes. The paper reports excellent rejection of knowledge below threshold, and positive transfer above threshold. In particular, an informal variant achieves strong transfer in this latter regime, and the paper discusses the factors which may influence the strength of this transfer.