Particle Filters as Weighted Samples of Hidden States
Summary
The document explains particle filtering through a comparison with the unscented Kalman filter. Both methods propagate representative states through a model and update them after receiving a measurement. A UKF uses a small collection of sigma points and covariance estimates, while a particle filter uses many randomly placed particles and assigns each a weight according to how well it matches the new observation.
After weighting, the particle filter resamples particles, favoring those with higher weights, and repeats prediction and measurement updates. The growing particle cloud is intended to approximate the distribution of the hidden state. A second answer offers a sampling analogy: importance sampling can focus samples on regions supported by evidence, whereas simple rejection sampling may miss an important region. These are conceptual explanations rather than a complete algorithmic specification; the document does not discuss proposal design, degeneracy, resampling choices, or convergence guarantees.
Key ideas
- A particle filter represents uncertainty about a hidden state with a collection of candidate particles.
- Particles are propagated through the model before new observations arrive.
- Measurement compatibility determines particle weights, and weighted resampling forms the next particle set.
- A UKF can be viewed intuitively as a sparse counterpart using a small number of sigma points.
- Importance sampling can concentrate samples in plausible regions, though the explanation omits implementation details and guarantees.
Tags
Full text
# Can someone explain the particle filter algorithm in detail with intuition # Can someone explain the particle filter algorithm in detail with intuition I am trying to understand particle filters and their application but i am not able to understand the underlying methodology. I have read a few sources but either the language is not clear or they dive into mathematics too quickly. I know kalman filters and have wrote a basic implementation in R myself after little help.I would like to have a similar understanding of particle filter. Edit : A reference to a very good paper/tutorial would also be fine. ## Answer by babelproofreader (score 1) https://quant.stackexchange.com/a/55209 > I know kalman filters and have ... If this knowledge extends to Unscented filters, UKF, you can think of the UKF being a sparse particle filter. With a UKF you have a few sigma points which are propagated forward via your model function and then after the measurement update these sigma points are updated via covariance estimation. With a particle filter, instead of a few sigma points you have very many more randomly allocated particles which are propagated forward via the model function and after the measurement update these particles are weighted according to their closeness to the new measurement. A new set of particles is then generated by weighted sampling of the "old" particles, and then we're ready to start the next round of propagation and update. Hopefully, over time, the cloud of particles converges to the true underlying state. A completely non mathematical explanation is given at https://www.youtube.com/watch?v=aUkBa1zMKv4. MATLAB/Octave code for this youtube video is available at http://www.it.uu.se/katalog/andsv164/Teaching/Material/PF.m ## Answer by K A Narayanankutty (score 0) https://quant.stackexchange.com/a/75208 Supposing you put random points over a surface, over which the Density function is a terrain(hills,valleys) i.e.,you objective is go to a place where some information about your target is in the form of a peak, the sampling gives you an intution of the location of the target. Then you want to reduce the area of your sampling space as you gather more information on the target. If you adopt rejection sampling, you may sometimes miss the target peak during sampling. So you adopt to a method called importance sampling. For this you choose an appropriate PDF(even uniform pdf) and use the ratio of the heights of your PDF with the new selected PDF covering your space of sampling. The process goes on as Predict based on information, and Update using this information. This is how Monte Carlo methods work.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.