This is a simple forward simulator of stochastic compartmental models meant to be used for comparison of Monte Carlo methods to approximate values of interest.
The architecture, data generation, and training setup seem OK for now. Sometimes, training yields good models with (validation and test) errors in the order of 10e-2 for small SIR instances.
What's left to do?
- Hyperparameter tuning
- Use the neural stochastic matrix to compute the expected number of infectious people at the moment of peak (observed empirically); to compute the expected end of the epidemic (and compare against estimator)
- Think about alternatives to mean absolute error as loss function
Point 2 is the main open problem now. How does one evaluate products of the matrix if we have a succinct representation thereof?