Machine learning meets individual-based modelling: Self-organising feature maps for the analysis of below-ground competition among plants

Authored by Uta Berger, Ronny Peters, Yue Lin

Date Published: 2016

DOI: 10.1016/j.ecolmodel.2015.10.014

Sponsors: No sponsors listed

Platforms: NetLogo

Model Documentation: Other Narrative Flow charts Mathematical description

Model Code URLs: Model code not found

Abstract

Individual-based models (IBM) simulate populations and communities whose dynamics are shaped by the properties, interactions and behaviour of the constituent organisms as well as the corresponding abiotic boundary conditions. Structurally realistic IBM can provide insights into the functioning of such systems and predict the effects of variable scenarios. We suggest complementing IBM with machine learning (ML) methods in order (i) to visualise correlation patterns between model inputs and model outputs, (ii) to provide simulation-based decision tools for non-modellers, and (iii) to derive information about factors difficult to obtain in the field on the basis of data that are more readily measurable. On top of this, ML methods can complement the established pattern-oriented modelling approach used to analyse the behaviour of IBM and to detect model uncertainties. As an example to demonstrate the strength of an IBM-ML connection, we combined the individual-based Plant Interaction Model (Pi model) with self organising feature maps (SOM) - a special type of ML. Based on simulation experiments with complete knowledge of the simulated system, the SOM was trained and used to visualise the nonlinear relationship between two IBM inputs (namely the mode of below-ground competition and below-ground resource limitation) and two model outputs (the mortality rate and the Clark Evans Index of the spatial distribution of plants). Our study also highlights an application of the SOM to infer the modes of below-ground competition (either symmetric or asymmetric) from the remaining measurable variables (resource limitation, mortality rate and Clark Evans Index). This procedure was successful in 92\% of cases, revealing its great potential as a means to assess parameters difficult to measure in nature. This example shows that SOM are powerful tools to revert the hierarchy of variables and to generalise dependencies of parameters in individual based modelling. (C) 2015 Elsevier B.V. All rights reserved.

Tags

behavior ecology pattern information growth Predict Abundance Artificial neural-networks