Is every SMC method a particle filter?

No. Particle filters address sequential state inference, while the broader family also includes other sequences of distributions.

What does effective sample size measure here?

The common weight-based diagnostic summarizes concentration of weights; it is not a complete measure of accuracy or particle diversity.

Must resampling happen after every observation?

No. Some methods resample adaptively, for example when a diagnostic crosses a chosen threshold. The policy is part of the algorithm design.

AI Glossary / Graphs and symbolic AI / Advanced

Sequential Monte Carlo

Definition

Sequential Monte Carlo, or SMC, is a family of methods that approximates a sequence of probability distributions using weighted random samples, often called particles. Particle filtering is a prominent application, but SMC also applies to sequences of targets that do not represent physical time.

What it means

Many inference problems require updating a probability distribution that cannot be calculated conveniently in exact form. A tracking system, for example, may need to represent several plausible positions rather than one best guess.

SMC represents such uncertainty through a collection of candidate values and weights. The particles are computational samples, not physical objects. Their weighted distribution approximates the target distribution, and averages over them approximate quantities of interest.

“Sequential” means that the method moves through a succession of targets, reusing information from previous stages. In a particle filter, new observations commonly define those stages. In other applications, intermediate distributions provide a bridge to a difficult final target. That broader scope is why SMC should not be defined solely as robot tracking.

How it works

A typical particle-filtering cycle propagates candidate states through a transition model, updates their weights using the latest evidence and, when appropriate, resamples. Resampling gives more descendants to particles with larger weights and fewer to particles with small weights.

In a bootstrap particle filter, propagation uses the transition distribution, allowing the observation likelihood to supply the new weighting factor. More general proposal distributions require an importance-weight correction; simply applying the likelihood is not a universal SMC recipe.

Over repeated updates, a few particles can carry nearly all the weight. Resampling addresses this weight concentration, but may reduce the number of distinct surviving candidates and introduce additional randomness. Proposal design, particle count, resampling choices and any valid rejuvenation moves all affect performance. A finite collection remains an approximation, and a poorly specified observation or transition model cannot be repaired merely by using more particles.

Three practical examples

A robot-localization system maintains several possible poses when a sensor reading is compatible with multiple places.
An equipment-monitoring model tracks uncertainty about a changing hidden operating condition as measurements arrive.
A Bayesian computation project uses a sequence of intermediate distributions to approximate a difficult posterior distribution for fixed, unknown parameters.

Think of it this way

Think of maintaining a crowd of candidate explanations. New evidence changes how much attention each explanation receives, and resampling redistributes the crowd toward better-supported regions.

The analogy is limited: particles are mathematical samples, not independent experts voting on truth. Duplicating a particle does not create new evidence, and a crowd that never explores an important possibility may still produce a misleading approximation.

Worked example

A small weighted update

Suppose three propagated particles represent positions 0, 1 and 2, each initially carrying weight one-third. In an invented observation model, the latest reading has likelihoods 0.20, 0.40 and 0.20 at those positions. The unnormalized weights are therefore 0.0667, 0.1333 and 0.0667, rounded.

After normalization, the weights are 0.25, 0.50 and 0.25. The estimated mean position is 0 × 0.25 + 1 × 0.50 + 2 × 0.25 = 1. This mean does not imply that the position is known exactly: uncertainty remains on either side.

A common weight-based effective-sample-size diagnostic is 1 divided by the sum of squared normalized weights. Here it is 1/(0.0625 + 0.25 + 0.0625), approximately 2.67. If four offspring are drawn by multinomial resampling, their expected counts are 1, 2 and 1, but an actual random draw can differ. Resampling then gives the offspring equal weights; it does not guarantee four distinct positions.

Illustrative case study

Fictional case study: Birch Tracking evaluates a particle filter for a small indoor vehicle. Two corridors initially produce similar observations, so the particle collection represents both possibilities.

An early implementation resamples aggressively and loses particles from the corridor the vehicle actually occupies. Later observations favor that corridor, but the remaining proposals seldom return there. Increasing confidence around the wrong location makes the display look stable without making it correct.

The team reviews its proposal mechanism, resampling policy and recovery behavior, and evaluates complete recorded routes with known reference positions. The fictional lesson is that particle count and a smooth plotted estimate are insufficient evidence of reliable tracking. Maintaining relevant alternatives and checking uncertainty against reality matter too.

Common misconceptions and mistakes

Every particle has equal influence. Weighted particles can contribute very differently before resampling.
Resampling adds independent observations. It reallocates computational effort using existing weights.
The particle mean describes every possible outcome. A mean can hide multiple separated regions of probability.
More particles fix a wrong model. They may reduce sampling error while preserving model error.
SMC and Markov chain Monte Carlo are identical. They are different families, although an SMC method can incorporate suitable MCMC moves.

Business use

SMC is useful when a business needs approximate uncertainty estimates for a changing or otherwise difficult inference problem. Its flexibility can accommodate models whose distributions are not adequately represented by one Gaussian approximation.

Evaluate latency, memory, repeated-run variability and calibration alongside average estimation error. Record the particle budget, proposal, observation model and resampling policy so that comparisons are meaningful. Rare but consequential scenarios deserve explicit evaluation; an attractive typical trajectory does not establish that the particle population reliably covers unusual states.

Check your understanding

After resampling, all 100 particles have equal weights, but 95 are copies of the same earlier particle. Does equal weighting prove good coverage of uncertainty?

Answer: No. Equal weights do not establish diverse support or accuracy. The propagation and proposal mechanisms, target distribution and evaluation evidence still matter.

Related terms

Learn next

Read importance sampling and particle filtering. Repeat the three-particle calculation with a likelihood concentrated at position 2. Compare the weighted mean, weight-based effective sample size and possible resampled offspring rather than treating any one number as a complete diagnosis.

Sources

Adam M. Johansen, Sequential Monte Carlo: Particle Filtering and Beyond: https://wrap.warwick.ac.uk/id/eprint/136101/ and its linked accepted manuscript. The overview and relevant weighting/resampling discussion were reviewed. Russell and Norvig's particle-filtering teaching slides provide a bootstrap-filter example: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter15a.pdf. Numerical examples and the tracking case above are original illustrations.

Previous: Sequential model-based optimization Next: Sequential Monte Carlo approximate Bayesian computation