Applying Machine Learning to Learn System Dynamics Models for Urban Systems

Abstract

System Dynamics (SD) is an approach to study the nonlinear behaviour of complex systems over time. SD models provide a high­level understanding of the system and aid in designing policies to achieve specific system behaviours. Conventional SD modelling requires an intensive amount of time, human resources and effort. Applying Machine Learning (ML) techniques benefits the modelling process in saving on resources. It also has the potential to provide insights into the system and prevent subjective­ ness of the modeller. This work proposes two methodologies, EvoNN and EvoESN, to learn SD models automatically for the urban system from observations under different levels of prior knowledge. EvoNN solves the automated equation formulation task for a Causal Link Diagram (CLD) and annotates it with Shallow Neural Networks (SNNs) as surrogate equations. The annotated CLD can be further used in simulating the system behaviour. We provide experimental results on a real­world urban system in Am­sterdam as well as the evaluation of the simulation results. The second methodology, EvoESN learns both the structure and the quantitative relations in the model without the prior knowledge about the structure. Trained using observation data, the EvoESN produces satisfactory results on the real­world urban system. We further incorporate the judgement from the domain expert to evaluate the learned model. Applied on a more complex system, EvoESN shows solid reliability and scalability to handle large datasets. Both EvoNN and EvoESN stand as promising supportive tools for SD modellers and remain robust even when lacking system observations.