Keywords
model reduction, data mining, self-organising map, nonlinear projection, multispecies competition
Start Date
1-7-2004 12:00 AM
Abstract
Complex models of environmental systems typically depend on a large amount of uncertain parameters. Therefore, they are often difficult to handle and do not provide an insight into effective modes of the underlying system’s dynamics. Unlike earlier analytical attempts to find more effective model representations, we present a new combination of methods that only relies on data generated by complex, process-based models. These methods are taken from the field of data-mining and enable the recognition of patterns in measured or modelled data by unsupervised learning strategies. As these methods do not directly lead to a better understanding of the systems’ driving processes, we suggest the linkage between pattern recognition and process identification by a multi-stage approach. In a first step, a large data-base was produced by a mechanistic model for species competition in a virtual ecosystem for a range of parameter settings. Using Vector Quantisation and nonlinear projection techniques such as Self-Organising Maps and nonlinear Principal Component Analysis, typical states of the complex model’s dynamics as well as major pathways connecting these states were then identified. The visualisation of the results points to the existence of nonlinear transformations of former model state variables and parameters to few effective variables. Effective variables built this way preserve most of the model’s dynamic behaviour, while they are nonetheless easier to use and require much less parameterisation effort.
Reduction of complex models using data-mining and nonlinear projection techniques
Complex models of environmental systems typically depend on a large amount of uncertain parameters. Therefore, they are often difficult to handle and do not provide an insight into effective modes of the underlying system’s dynamics. Unlike earlier analytical attempts to find more effective model representations, we present a new combination of methods that only relies on data generated by complex, process-based models. These methods are taken from the field of data-mining and enable the recognition of patterns in measured or modelled data by unsupervised learning strategies. As these methods do not directly lead to a better understanding of the systems’ driving processes, we suggest the linkage between pattern recognition and process identification by a multi-stage approach. In a first step, a large data-base was produced by a mechanistic model for species competition in a virtual ecosystem for a range of parameter settings. Using Vector Quantisation and nonlinear projection techniques such as Self-Organising Maps and nonlinear Principal Component Analysis, typical states of the complex model’s dynamics as well as major pathways connecting these states were then identified. The visualisation of the results points to the existence of nonlinear transformations of former model state variables and parameters to few effective variables. Effective variables built this way preserve most of the model’s dynamic behaviour, while they are nonetheless easier to use and require much less parameterisation effort.
