Folie 1

1
A FEASIBILITY STUDY PARTICLE FILTERS FOR MOBILE
STATION RECEIVERS
CSNDSP 2006
Infineon
Michael Lunglmayr, Martin Krueger, Mario Huemer
2
Contents

• Introduction
• Simulation Model
• Particle Filters
• Particle Filters for Equalization
• Simulation Results

3
Introduction

• Particle Filters popular in e.g. image
  recognition, positioning,...
• Aim of this work Equalization with particle
  filters
• ?Symbol estimation for GSM/EDGE in a multipath
  propagation environment

4
Simulation Model
5
Simulation Model
6
Particle Filters

• Connection to Equalization Estimate p(xkyk)
  and choose those state with the highest
  probability
• Straight Forward Method calculate p(xkyk) for
  every state
• ? Effort to high for practical systems

7
Particle Filters

• Connection to Equalization Estimate p(xkyk)
  and choose those state with the highest
  probability
• Straight Forward Method calculate p(xkyk) for
  every state
• ? Effort to high for practical systems
• Importance SamplingPrinciple If p(xkyk) would
  be known, it could be sampled? Particles
  then for N??

8
Particle Filters

• Bad News p(xkyk) is not known because it is to
  be estimated!
• But If we can sample a different probability
  functionq(xkxk-1,yk) (importance sampling
  function) and weight the particles with an
  importance weight

9
Particle Filters

• Bad News p(xkyk) is not known because it is to
  be estimated!
• But If we can sample a different probability
  functionq(xkxk-1,yk) (importance sampling
  function) and weight the particles with an
  importance weight
• Example q(xkxk-1,yk) p(xkxk-1)?

10
PF for Equalization

• Probability functions for GSM/EDGE

11
PF for Equalization

• Probability functions for GSM/EDGE
• Until now Sequential Importance Sampling
  (SIS)?But not very efficient yet!

12
Resampling
13
Particle Filter Algorithm
14
Implementation
15
Simulation Results
GMSK
16
Simulation Results
17
Conclusion

• Particle Filters can outperform existing
  algorithms
• Disadvantage computational complexity
• But complexity depends only linearly on channel
  length? e.g. Promising use in extremely
  broadband communication systems with long
  impulse responses