SDE inference course

  • 12 November at 15.15-17.00 (MVL14): Lecture (Umberto). Slides are in this Box folder. There are 2 versions: one named "static" with disabled transparencies and animations, for easy browsing. Then one named "dynamic", only useful to enable the particle filter animation (and transparencies).

Suggestions for further reading:

1. For an easy intro to the bootstrap filter and SMC, see Umberto's blogpost .

2. Chapter 7 in Särkkä https://users.aalto.fi/~ssarkka/pub/cup_book_online_20131111.pdf .

3. SMC/particle filters can well be used beyond likelihood estimation. The available literature is, quite frankly, overwhelming in size. However why not looking into what is being done by our Swedish colleagues. Here is a quite large report by the excellent Naesseth/Lindsten/Schön https://arxiv.org/abs/1903.04797


  • 16 November at 13.15-15.00 (perhaps 60 min only) (MVL14): Lecture (Umberto). Zoom details have been sent via email.

Slides are in this Box folder.

For another digestible (I hope!) intro see my blog post and references therein. References in the post also point to papers considering the tuning of pseudomarginal algorithms.

The original papers showing the pseudomarginal property of being an "exact-approximate" method are Beaumont (2003) and Andrieu and Roberts (2009) where the latter studied the problem with higher mathematical detail. It ain't a walk in the park: you may want to look at Dahlin and Schön 2017 for a more friendly paper, discussing numerics and implementations in R.

If you wish to consider simultaneous (exact) inference for states and parameters, ie sampling from $\pi(\theta,x|y)$, you can use the algorithm denoted PMMH from Andrieu, Doucet and Holenstein (2010) and explained in Wilkinson's blog.


  • 23 November 13.15-15.00 (MVL14) : Students presentations.

Slides of the presentations are in this Box folder.

  • 30 November 13.15-15.00 (MVL14): Lecture (Moritz) following the introduction chapter of Moritz' thesis + students presentations.

  • 8 December 15.15-17.00 (MVL14): Students presentations

  • 16 December, 15.15-17.00: Axel's presentation on Partially Observable Markov Decision Processes (POMDP)

  • 16 June final presentations


Addenda:

https://paschermayr.github.io/category/state-space-models/ is a not so bad discussion of state space models Julia code