• Bit more writing on comment reply. Sent to Alex and Alan for feedback.


Demonstrating robustness of the GP approach is remarkably difficult ironically because the comparison methods are not particularly robust; most often due to poor initial conditions. Doing all I can to make the MLE and parametric Bayesian cases suitably automated such that the can give reasonable performance as I loop over various simulation parameters and models without much hand-holding of each of the estimators.

Also an automatic sensitivity analysis gets pretty computationally intensive and generates a lot of different outputs to keep track of, trying to improve this through cleaner code, parallelization, caching, chunk dependencies, and namespaces for scripts.

  • Playing with mcmc convergence over initial conditions
  • Running case without measurement error for parametric bayesian models
  • Adaptuing runs for parallel and for execution on farm cluster
  • Consider ADMB for the MLE estimate to be a bit more robust for automated cases.