Abstract for: Model learning using genetic programming under full and partial system information conditions

Modelling is often an expensive and time-consuming process. Issues with modelling efficiency and implications to model adoption are long recognized in the broad modelling/simulation as well as system dynamics literature. Computational methods for supporting model learning provide an opportunity for addressing these limitations by providing computational tools that have the capabilities to analyse large datasets and explore possible model structures and parameter that can explain the behaviour of interest. This paper, a part of an ongoing research, aims to contribute to the area of using machine learning to support learning about system dynamics models. Towards this aim, the paper has two objectives. First, we present and describe the proposed methodology which use genetic programming method under the assumptions of full and partial information available about the system. Secondly, we present three case studies to develop and test the proposed methodology. Preliminary results show that the genetic programming was able to find good approximated models for first and second order system dynamics model. However, the results for third order models the learned models were not good approximations for the original models. Combining genetic programming and embedding reconstruction technique showed better results in the case of third order model.