Limitations of the discrete Bayes filter

Parent:  Discrete Bayesian filter
Source: rlabbe

Limitations of the discrete Bayes filter

• Scaling
  • Dog tracking example is one-dimensional, but in real life we often want to track more things (e.g. 2D coordinates, velocities)
  • Multidimensional case: store probabilities in a grid
  • 4 tracked variables: O(n^4) per time step
  • High computational cost with high dimensionality
• Filter is discrete and therefore gives discrete output
  • But a lot of applications require continuous output
  • Discretising a solution space can lead to lots of data (depending on accuracy required) –> calculations for lots of different probabilities!
• Filter is multimodal
  • Not always a problem, e.g. particle filters are multimodal
  • But not always a good things either, sometimes not a realistic representation of the reality (e.g. 40% in this location, 30% in the other location)
• Requires measurement of the change in state (dog tracking example assumes movement by 1 unit)

We want a unimodal, continuous filter –> achieve this by using Gaussian distributions