At around 12:15 you say, given we know x_{t-1} is exactly that point and we guess that at time t it could not have gone that far and randomly sample x_t at the upper left point that you show. Could we control action information here? Then, in Step 2: Reweighting you say we end up with a set of new point. Are these the new points (assuming they're different from the previous ones) that you sample based on X_{t-1}. Do you take the point X_{t-1} and set the new points around it as these are all the points where the state could have transitioned to?
@oldcowbb4 жыл бұрын
so the benefit of particle filter is to simply avoid calculating PDFs?
@devonk2985 жыл бұрын
can an object be a moving price in a time-series?
@berty385 жыл бұрын
Definitely. That’s a very common usage of time series analysis, though I’m a bit skeptical that people are using particle-based Markov models for those these days.
@peaelle423 жыл бұрын
15:56 oh really? you get arbitrarily close to the true distribution just by spamming points? what happens if you have noisy measurements that screw up your reweighing?
@shrinivasiyengar57994 жыл бұрын
I understand you arriving at p(Y_t = y, X_t | Y_{1 : t-1)} - (1). I also understand p(Y_t = y | Y_{1 : t-1} - (2) . What I do not understand is how you got the posterior p(X_t | Y_{1 : t-1}) equals (1) divided by (2). Am I missing some probability identity here?
@captainbig5345 Жыл бұрын
that‘s also my question
@captainbig5345 Жыл бұрын
oh i know, just using this joint distribution formula p(a,b|c)=p(a|c)p(b|a,c)