Georgiy V. Bobashev, Stephen P. Ellner, Barbara A. Bailey
Predictive models based on past data could be good predictors of the future outcomes; however, they usually don't explain the causal and feedback relationships leading to the outcome. Conversely, mechanistic models could uncover complex interaction between underlying processes, but sometimes their calibration and validation could be unrealistic. Combining the two approaches into a semi-mechanistic model can lead to a winning combination. We present examples of historic epidemic data as well as simulated data, where a combination of neural networks with a mechanistic Susceptible, Exposed, Infected and Recovered (SEIR) model produces more reliable predictions with less parameterization.