S. Bosse, Surrogate Predictive and Multi-domain Modelling of Complex Systems by fusion of Agent-based Simulation, Cellular Automata, and Machine Learning, Proc. of the SIMUL 2021 Conference, The Thirteenth International Conference on Advances in System Simulation, IARIA,
3. - 7. October 2021, Barcelona, Spain, & Online
Modelling of complex dynamic systems like pandemic outbreaks or traffic flows in cities on macro-level is difficult due to a high variance on entity micro-level and unknown or incomplete interaction models. Agent-based and Cellular Automata (CA) simulations based on micro-level modelling can be used to investigate the outcome of system observables in a sandbox. For a reasonable accuracy a high number of agents, sufficient behaviour variance, high computational times, and calibrated model parameters are required. Surrogate predictive modelling of the multi-agent system can be used to replace time-consuming simulations. In this work we present a hybrid aprroach combining Agent-based Simulation, probabilistic contextual CA, and Machine Learning (ML). We investigate the replacement of the ABS-CA by surrogate ML models trained by simulation data. The predictive model is state-based and applied to time-series data to predict future development of aggregated system observables. We discuss and show the negative impact of uncalibrated real-world sensor data on time-series prediction and an improvement by surrogate modelling of simulation. A use-case of pandemic simulation using real-world statistical data is used to investigate and evaluate the suitability and accuracy of the proposed methods and to show the high sensitivity of surrogate modelling on distorted and biased data.