Views: 4 | Downloads: 0
In this thesis, we propose a novel methodology for learning ensembles of process-based models,
which addresses the task of automated modeling of dynamic systems. Mathematical
modeling is at the heart of science and engineering and is often employed for representing
and reasoning about the complex reality. Scientists and engineers construct complex
mathematical models to study, control and/or predict the behavior of real-world systems
under various conditions. Constructing mathematical models, however, is a nontrivial and
demanding task that requires an extensive expert knowledge and empirical observations of
the system at hand.
There are two main paradigms for modeling dynamic systems: theoretical (knowledgedriven)
modeling and empirical (data-driven) modeling. In the former, a domain expert
first derives a proper structure of the model, based on extensive knowledge about the system
at hand. Measured data can then be used to estimate the (constant) parameters in the
model, which, however, can also be set manually based on domain knowledge. The latter
uses measured data in a trial and error process to search for such a combination of model
structure/and parameters that best fits the observed behavior. Process-based modeling
aims at integrating the theoretical and the empirical paradigms for modeling dynamic systems.
More specifically, it aims at combining high-level process-based modeling knowledge
and observed data from the system at hand. These are used for learning process-based
models (PBMs) where a PBM is a detailed, interpretable and modular representation
of a dynamic system. The state-of-the-art process-based modeling approaches have been
successfully applied to a variety of modeling tasks in a number of real-world domains. They
have been predominantly used for describing the observed behavior of the modeled system.
While focusing on the provision of detailed and accurate descriptions of the observed
system, these approaches have only shown limited predictive power when applied to tasks
of forecasting future system behavior.
To address this limitation of the process-based modeling paradigm, we focus on a well
established approach for improving the predictive performance of models in machine learning
of the paradigm of ensemble learning. Ensembles are a machine learning paradigm
leading to accurate and robust models and are predominantly applied to predictive modeling
tasks. Ensemble models comprise a finite set of predictive models, whose combined
output is expected to yield an improved predictive performance as compared to the predictive
performance of an individual process-based model.
The proposed methodology for learning ensembles of process-based models of the main
contribution of this thesis combines two lines of research. First, it extends the scope of
existing approaches to learning process-based models of dynamic systems, predominantly
employed in a descriptive modeling setting, towards predictive modeling. Second, it follows
the basic principles of ensemble learning, as is common in machine learning, and translates
them into a methodology for modeling dynamic systems. More specifically, we focus on a
methodology for learning homogeneous ensembles of process-based models, which consist of models constructed by using the same machine learning algorithm on different samples
of the training data and/or knowledge.
The proposed methodology includes algorithms for learning ensembles of process-based
models by (1) bagging, i.e., learning from bootstrap samples of the data, (2) boosting,
i.e., learning from error-weighted samples of the data, (3) learning from random library
subsamples and (4) a combination of bagging and learning from random library subsamples.
Furthermore, the methodology includes methods for simulating ensembles of process-based
models that adapt the standard methods for combining predictions of ensemble constituents
in machine learning. Overall, the proposed methodology provides a general and modular
framework that allows for adapting other ensemble methods from machine learning to the
particular context of learning process-based models.
We conduct an extensive experimental evaluation based on a novel empirical evaluation
framework for process-based models of dynamic systems proposed in this thesis. This
framework is employed for estimating the predictive performance of process-based models
and ensembles thereof. This evaluation framework also identities the optimal design choices
within the proposed different ensemble methods. Furthermore, it allows for investigating
whether the diversity among the individual ensemble constituents is related to the predictive
performance of the ensemble. Finally, we analyze the computational efficiency of the
proposed methods. These analyses relate to both the empirical evidence gathered from
the performed experiments, as well as a theoretical study of the individual components in
each of the proposed ensemble methods.
We apply the proposed methodology and evaluate its performance on a set of problems
of automated predictive modeling of population dynamics in real-world aquatic ecosystems.
The experimental results provide important evidence that ensembles of process-based models,
in general, yield significantly better predictive performance compared to an individual
process-based model. While the different proposed ensemble methods have similar predictive
performance, the analysis of their computational efficiency reveals that ensembles
learned from random library subsamples are the most efficient to learn. They are constructed
using time comparable to the time needed to learn a single process-based model.