nchopin/particles

Forum Creation

Closed this issue · 3 comments

afcarzero1 commented

I am very interested in this topic and I have been exploring the library in the past months in my free time. I find it hard to find people knowledgeable in the topic. Maybe a forum for questions might be created.

I am asking anyway my question here in the meantime:

As far as I understand all particle filters rely on the assumption that we have a model (even vague) of the underlying process. Has it ever been tried to learn the model directly from data and then apply the particle filters? I have a very complicated process with many observations and I think particle filters are the answer. But creating a model is basically imposible.

nchopin commented

hi,
there is a "discussions" thing on github, I could activate it if you want, that seems like a better place than "issues" (which is for tracking bugs).

Regarding your question : as far as I know, there have been few attempts in this direction; one way to go would be to design a partly (or totally) nonparametric state-space models based on DPMs (Dirichlet Process mixtures); see this for instance :
https://arxiv.org/abs/1810.09291
or this (sorry, this is not on arxiv):
https://www.sciencedirect.com/science/article/pii/S0165168416000633?casa_token=xjtRJnI6BC0AAAAA:n4GqkbvamIrV853NBxA1SLYjOCCJ3sJ6_R0fTzZxWgxt7DK47wHrCOQEqQcamZGpmSC7W_wP4A

I am a bit sceptical about such approaches, because the estimation of parametric state-space models is already quite challenging; i.e. even with lot of data, some of your parameters may be hardly identified. So trying to use non-parametric model looks too ambitious. But there might be practical scenarios where it actually makes sense.

nchopin commented

Closing this since the OP did not reply, but happy to activate the discussions tab if anyone thinks it could be helpful.

lukego commented

learn the model directly from data and then apply the particle filters?

I am also interested in this problem.

This paper might be interesting to you: https://arxiv.org/abs/2307.09607. They treat it as model selection by particle filtering over a large model space.

  • 👍2