Quantised causal inference based on directed information
Project Leader: Girish Nair
Staff: Erik Weyer
Primary Contact: Girish Nair (email@example.com)
Keywords: complex systems; knowledge discovery; machine learning; signal processing; system identification
Disciplines: Electrical & Electronic Engineering
Domains: Networks and data in society, Optimisation of resources and infrastructure
In the field of causal inference, the aim is to determine the causality relationships between 2 or more observed time series, under suitable assumptions on the underlying generating model classes. Causality here is defined in terms of conditional probabilities, based on Granger’s ideas [e.g. Amblard and Michel, Entropy, 2013].
All current techniques for causal inference are based on full-resolution measurements of the time series of interest. However, in many remote sensing applications only quantised, low-resolution measurements may be available.
This introduces several major theoretical challenges not addressed in the existing literature.Firstly, existing causality inference/system identification techniques cannot be applied directly. Secondly, optimal quantiser design becomes a critical issue. This project will explore these challenges, by using the Marko-Massey directed information and/or the nonstochastic directed information [Nair, IEEE CDC 2012; Nair, IEEE CDC 2015] as measures of causality.