Date: February 6, 2024

Topic: Terminology

Recall

Filtering, genetic algorithm and particle filters

Notes

Terminology

Filtering

Process noisy data (observations / motions) in a way that aims to estimate the underlying true value of a system

Genetic Algorithm

A class of iterative algorithms which evolves a population of candidate hypotheses

Components:

• Starting population
• Fitness function (score, weight)
• Selection process (offspring)
• Crossover
• Mutation

Particle Filters

A genetic algorithm that represents its belief distribution using a population of hypotheses

Particle filters need to balance between speed and accuracy

Approximation (Beliefs)

• In order to perfectly represent a distribution, we need to make infinitely many samples
• The more samples we have, the slower it takes to process
• Hence there is a trade-off between accuracy and speed in particle filters

Particle filters have an iterative loop where good predictions are kept and bad ones are filtered out continuously

Task

To localize a target (real time) given map, ground truth, target observation, target movements

• Need to be fast enough based on demand (e.g., if we need to know every second)

Localization Loop

1. Evaluating a set of hypotheses against a measurement received from the target
2. Keep the good mones
3. Copy the actions that the target performs
4. Diversify into the set of hypotheses
5. Derive a best guess at the target’s location based on the current set of hypotheses
6. Repeat from Step 1

<aside> 📌 SUMMARY: Particle filters are used for localization by instantiating many particles. These particles are continuously updated and filtered out to give the best guess of the robot

</aside>

Date: February 6, 2024

Topic: Components

Recall

Components of a particle filter

Notes

Components:

1. Initialization

2. Weighting

3. Resampling

4. Fuzzing

5. Mimicking

6. Aggregating

Initialization

Process of defining the initial set of hypotheses

• Coverage (sparse/dense)
• Prior knowledge