High dimensional sampling kickoff meeting

Wednesday 10 Oct 2018, 20:00 → 22:00 Australia/Adelaide

Other Institutes

We discussed possible techniques to use for high dimensional sampling problems. Our rough brainstorm included:

Bayesian optimisation
Active learning
Multi-fidelity optimisation
Likelihood-free inference? (arXiv:1808.00973) we should invite Gilles to a future meeting.
Multiarmed bandit, Thompson sampling.
Revisiting BAMBI idea with a better machine learning module (plus see 1708.06615)
Training classifiers to get quick estimates of which models are physical for complex cases (e.g. composite Higgs, 19D SUSY).

Joaquin has promised to send round some relevant literature and most of our homework until the next meeting will be digesting this literature.

We will start by comparing these approaches on nD Gaussian functions, and in particular we want to try functions with lots of zero likelihood regions, and isolated peaks (this is intended to mimic cases where most of a model is unphysical, e.g. SUSY/composite Higgs). We need to define how we are going to compare the efficacy of each method (e.g. how well does it find the best fit point? How well does it map the 2 sigma and 3 sigma profile likelihood contours? If applicable, how well does it map the posterior?). Csaba will send some relevant information, and Judita will look into this.

Later, we will define a suitable physics likelihood for applying the techniques.

We decided to use a custom Python interface (developed on github) for the toy, analytic examples, and to switch to using GAMBIT once we get to the point of running physics examples.

Roberto will look into setting up a mailing list and slack channel.

Martin will send a new doodle poll for a meeting time now that the clocks are changing. The suggested frequency is once every 2 weeks in the initial stages whilst we are getting things up and running, on the promise that we drop meetings if there is nothing to discuss.

 

20:00 → 20:20 Introduction

Speaker: Martin John White (University of Adelaide (AU))

DarkMachinesKickOff.pdf

20:20 → 22:00 Discussion