Download files
Abstract
Complex systems are ubiquitous in both natural and man-made systems. Understanding their behaviors is not a trivial task and modeling techniques are often used to analyze and experiment with their behavioral changes. Most end-to-end approaches for modeling complex systems, such as system dynamics, involve several processes, many of which rely heavily on the expertise of human modelers. In this context, the key focus of this research is to improve the performance of system dynamics processes by developing computational methods.
This thesis offers three main contributions to this field of research. Firstly, an echo-state network-based technique for learning the causal loop diagrams is proposed. Its central idea is to encode an echo-state network's dynamic reservoir with a known number of nodes, equal to the number of key system variables identified, and then train the network using the system observations to match the observed behavior.
Secondly, a novel genetic programming-based symbolic regression ensemble method based on pre-defined causal relationships between system variables is applied to learn the system equations. Information about these relationships is used to decompose the problem space. The ensemble members independently learn the equations for different output variables, with these learned models then combined to generate the final model.
Finally, an integrated system for supporting the modeling of system dynamics which facilitates data-driven learning of the different processes involved, including causal loop, and stock and flow diagrams, equations and the values of the model parameters using multiple computational intelligence techniques, is presented. A prototype for the support system is developed to consist of two main components: a graphical user interface that allows the modeler to interact with the tool; and the core part of the support system, a learning engine, which is the back-end of the system, comprises the data and model repositories, and implements different intelligence algorithms.
Although the actual utility of these methodologies can only be known through their use by modelers of system dynamics, we conduct a number of experiments on several real case studies to demonstrate their performances. The empirical results verify their efficiency in terms of learning models similar to the target ones.