Learning from multi-level behaviours in agent-based simulations: a Systems Biology application

Abstract

This paper presents a novel approach towards showing how specific emergent multi-level behaviours in agent-based simulations (ABSs) can be quantified and used as the basis for inferring predictive models. First, we first show how behaviours at different levels can be specified and detected in a simulation using the complex event formalism. We then apply partial least squares regression to frequencies of these behaviours to infer models predicting the global behaviour of the system from lower-level behaviours. By comparing the mean predictive errors of models learned from different subsets of behavioural frequencies, we are also able to determine the relative importance of different types of behaviour and different resolutions. These methods are applied to ABSs of a novel agent-based model of cancer in the colonic crypt, with tumorigenesis as the global behaviour we wish to predict.

Keywords:

• behaviour
• learning
• regression
• simulation
• systems
• system dynamics

Acknowledgements

Chih-Chun Chen acknowledges the advisory support of Christopher D. Clack and Sylvia B. Nagl as part of a multi-disciplinary collaboration between the departments of Computer Science and Oncology at UCL. David R. Hardoon acknowledges financial support from the EPSRC project Le Strum, (http://www.lestrum.org.) EP-D063612-1 and from the EU project PinView, (http://www.pineview.eu.) FP7-216529.

Notes

1 The Singular Value Decomposition (SVD) is a widely used technique to decompose a matrix into several component matrices, exposing properties of the original matrix. Using the SVD, we can determine the rank of matrix, quantify the sensitivity of a linear system to numerical error, or obtain an optimal lower-rank approximation to the matrix.