Recent Trends in Modelling and Simulation with Machine Learning

In recent years, the increased accessibility of large computing power, together with the availability of big data and the relevant progress in algorithms, led to an exponential increment of Machine Learning (ML) applications to predictive tasks concerning complex systems. In general, ML automatically builds a model of the system under study by exploiting an available large dataset of input features coupled to the corresponding expected outputs. While automatically learning models from data is an extremely powerful approach, in the case of complex systems' dynamics the generalization ability of ML models can easily be limited, that is, the predictions can be inaccurate once the model is applied beyond the bounds of previously observed data. On the other hand, a traditional field that deals with predicting the dynamics of complex systems is represented by Modelling and Simulation (M&S), in which, thanks to the knowledge of sufficient details concerning interactions and processes within the system under study, an abstract representation of it is developed and used for artificially reproducing its behaviour. Compared to ML models, when the M&S approach can be based on a sound theory about the functioning of the system, it can enable, in the line of principle, a more reliable forecasting of the system's dynamics beyond the bound defined by the observed historical behaviour. In spite of the outlined conceptual distance between the ML and M&S approaches, a recent and promising research trend encompasses their synergic combination for producing better data-driven simulation models. On the basis of an analysis of the relevant and recent scientific literature, the paper discusses new ideas and directions concerning M&S approaches based on advanced ML techniques.