Behavioural models in sequential inference about spatio-temporal structures

Project Summary: The project will involve new models and scalable inference methods for dynamic learning of behaviours and intentionality in multiple interacting objects, e.g. animals hunting prey, sea mammals/fish interacting in schools/shoals, aircraft or birds interacting in formation and high frequency financial modelling of multiple order books.
We will provide new methodologies for scientists and zoologists wishing to test/develop hypotheses about behavioural interactions, new methods for improved situational awareness/intent prediction in general tracking applications (civilian or military), and scalable inference in larger scale problems than currently manageable. The key objective of the project is to improve upon current state-of-the-art tracking algorithms (particle filters) for analysing interactions between multiple objects. One particular problem of interest is to reduce the complexity of the algorithm model such that large groups of interacting objects can be analysed with limited computational cost, where current methods are too slow.

The methodology will use multiple dimensional stochastic processes in continuous time. The underlying principles will be Bayesian updating of dynamically evolving objects, implemented using novel scalable combinations of Sequential Monte Carlo, message passing and machine learning algorithms. New probabilistic models of behavioural interactions will be employed within the existing tracking framework to incorporate inference of characteristics such as intention, future behaviour, health and leadership.

This project has applications in many domains of national and international importance, including the monitoring of wildlife (changes in arctic seal behaviours/ populations, sea mammals, fish), monitoring of large scale spatio-temporal effects in climate change, enhanced situational awareness in defense applications.