## likelihood-free Bayesian design [SimStat 2019 discussion]

*Related*

This entry was posted on September 5, 2019 at 10:50 am and is filed under Statistics with tags ABC, Austria, Bayesian design, discussion, dynamical systems, likelihood-free methods, ODEs, optimal design, Salzburg, SimStat2019, slides. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.

### 2 Responses to “likelihood-free Bayesian design [SimStat 2019 discussion]”

### Leave a Reply Cancel reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

September 5, 2019 at 5:59 pm

I looked up some of the references in your presentation (thanks!) and now see that the Kleinegesse and Gutmann reference is about an emulator-based approach to experimental design, citing earlier Gutmann work on emulators. I don’t know the literature here—is this the best thing to read and cite for emulation-based approaches to intractable likelihoods?

September 5, 2019 at 4:35 pm

I wrote a case study on the Lotka-Volterra model in Stan, simply because it seems to be the hello-world of ODE models in applied math courses. Stan can compute both the log likelihood and derivatives of an ODE system where the user only has to code the ODE. Our autodiff system does the heavy lifting by internally creating a coupled ODE with sensitivites so that we can control error in derivatives. The full technique’s in our code and also described in our arXiv paper on Stan’s autodiff. Our ODE devs (Yi Zhang is leading the effort now) have done a lot more with stiff solvers and efficient internal Jacobians than when I wrote the arXiv paper, but it describes the basic technique that’ll work fine for Lotka-Volterra. Yi’s even got some prototype PDE solvers working with sensitivities.

On another related topic, how do you feel about using emulation when it’s possible to draw from the sampling distribution (the likelihood viewed as a distribution with fixed parametes)? The idea is to fit a GP or neural network or some other non-parametric model of the intractable density function using simulations as data. I’m about to set off on writing a case study on that in Stan as people have been using it for complicated PDE models that are just a component of a larger joint model. But I haven’t seen much mention of the technique being used in the wild. The best use case I know is Maggie Lieu’s model of galactic mass using gravitationally lensed data.