Reducing computational costs of agent-based modeling of respiratory infection spread using a machine learning-based surrogate model

Agent-Based Models (ABMs) have proven to be an effective tool for describing and predicting the dynamics of respiratory infections and forecasting future outbreaks and have helped health organizations control the disease by developing effective intervention strategy. The use of ABMs is accompanied by very high computational cost, which limits their use in real time. Replacing ABMs with machine learning-based models that can replicate the output or couple the two models together is a solution to the computational cost problem. This paper proposes a machine learningbased surrogate model to simulate an ABM simulating the spread of respiratory infection in Saint Petersburg to reduce simulation time and maintain equivalent accuracy in estimates. The research was based on evaluating the performance of a set of machine learning models under different approaches as surrogate models to use in place of ABM. Methods for generating ABM output chains were compared and evaluated through experiments using single-model approaches or ensemble approaches as a predictive model for each time step in the output (independent multi-output and regression chaining) or hybrid models between agent-based and machine learning. The results indicated that there are several models capable of replicating the simulation output sequence of the ABM with a slight superiority of eXtreme Gradient Boosting within the regression chaining approach. In the hybrid approach, the Long Short Term Memory model with the first values of the output sequence within the feature space outperformed the other models in obtaining more accurate results and achieved the lowest Mean Absolute Error and Root Mean Square Error.