10
Rao-Blackwellised Particle Fil tering for Dynamic Bayesian Ne twork Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Embed Size (px)

DESCRIPTION

Problem Formulation  general state space model/DBN with hidden variables and observed variables.  is a Markov process of initial distribution  Transition equation  Observations  Estimate  recursion  not analytically, numerical approximation scheme

Citation preview

Page 1: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network

Arnaud DoucetNando de Freitas

Kevin MurphyStuart Russell

Page 2: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Introduction

Sampling in high dimension Model has tractable substructure

Analytically marginalized out, conditional on certain other nodes being imputed

Using Kalman filter, HMM filter, junction tree algorithm for general DBNs

Reduce size of the space over we need to sample Rao-Blackwellisation

marginalize out some of the variables

Page 3: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Problem Formulation

general state space model/DBN with hidden variables and observed variables . is a Markov process of initial distribution Transition equation Observations

Estimate recursion

not analytically, numerical approximation scheme

tzty

tz )( 0zp)|( 1tt zzp

},,{ 1:1 tt yyy )|( :1:0 tt yzp

)|()|()|()|()|(

1:1

11:11:0:1:0

tt

tttttttt yyp

zzpzypyzpyzp

Page 4: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Cont’d

Divide hidden variables into two groups and

Conditional posterior distribution is analytically tractable.

Focus on estimating (reduced dimension) Decomposition of posterior from chain rule

Marginal distribution

tz tr tx)|()|()|( 11,;11 tttttttt rrpxrxpzzp

)|()|()|( 11,;11 tttttttt rrpxrxpzzp)|( :1:0 tt yrp

)|(),|()|,( :1:0:0:1:0:1:0:0 tttttttt yrpryxpyxrp

)|()|()|()|(

)|(1:1

1:11:01:0,1:1:1:0

tt

ttttttttt yyp

yrprrpryypyrp

Page 5: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Importance sampling and RAO-Blackwellisation

Sample N i.i.d. random samples(particles) according to Empirical estimate

Expected value of any function of hidden variables

},,1);,{( )(:0

)(:0 Nixr i

tit )|,( :1:0:0 ttt yxrp

N

ittxrtttN dxdr

Nyxrp i

tit

1:0:0),(:1:0:0 )(1)|,( )(

:0)(

:0

)( tfI

),(1

)|,(),()(

)(:0

)(:0

1

:0:0:1:0:0:0:0

it

it

N

it

tttttNttttN

xrfN

dxdryxrpxrffI

Page 6: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Cont’d

Strong law of large numbers converges almost surely towards as

Central limit theorem

)( tN fI )( tN fI N

),0()]()([ 2tfNttN fIfIN

Page 7: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Importance Sampling

impossible to sample efficiently from target Importance distribution q

Easy to sample p>0 implies q>0

)),(()),(),((

)(:0:0)|,(

:0:0:0:0)|,(

:1:0:0

:1:0:0

ttyxrq

tttttyxrqt xrwE

xrwxrfEfI

ttt

ttt

)|,()|,(),(,

:1:0:0

:1:0:0:0:0

ttt

ttttt yxrq

yxrpxrwwhere

),(~

)),((

)),(),((

)(ˆ)(ˆ

)(ˆ

)(:0

)(:0

1

)(:0

1

)(:0

)(:0

1

)(:0

)(:0

)(:0

)(:0

1

11

it

itt

N

i

it

N

i

it

it

N

i

it

it

it

itt

tN

tNtN

xrfw

xrw

xrwxrf

fBfAfI

Page 8: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Cont’d

The case where one can marginalize out analytically, propose alternative estimate for

Alternative importance sampling estimate of

To reach a given precision, will require a reduced number N of samples over

tx :0

)( tfI

)( tfI

N

i

it

N

i

itt

ittryxp

tN

tNtN

rw

rwxrfE

fBfAfI

ittt

1

)(:0

1

)(:0:0

)(:0),|(

2

22

)((

)(),((

)(ˆ)(ˆ

)(ˆ )(:0:1:0

tttttt

tt

ttt

dxyxrqyrq

yrqyrprwwhere

:0:1:0:0:1:0

:1:0

:1:0:0

)|,()|(

)|()|()(

)(ˆ2tN fI

)(ˆ1tN fI

Page 9: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Rao-Blackwellised particle filters

Sequential importance sampling step For i=1,…,N sample:

and set:

For i=1,…,N evaluate importance weights up to a normalizing constant:

For i=1,…,N normalize importance weights:

Selection step multiply/suppress samples with high/low importance weights to obtain random samples approximately distributed

MCMC step Apply a markov transition kernel with invariant distribution given by

to obtain to obtain

),|(~)ˆ( :1)(

1:0)(

titt

it yrrqr

)ˆ( )(:0itr

)ˆ|ˆ()ˆ( )(1:0

)()(:0

it

it

it rrqr

)|ˆ(),|ˆ()|ˆ(

1:1)(

1:0:1)(

1:0)(

:1)(

:0)(

t

itt

it

it

titi

t yrpyrrqyrpw

1

1

)()()( ][~

N

j

jt

it

it www

)(~ itw

)~( )(:0itr )|~( :1

)(:0 tit yrp

)|( :1)(

:0 tit yrp )( )(

:0itr

Page 10: Rao-Blackwellised Particle Filtering for Dynamic Bayesian Network Arnaud Doucet Nando de Freitas Kevin Murphy Stuart Russell

Robot localization and MAP building

Problem of concurrent localization and map learning Location Color of each grid cell Observation

Basic idea of algorithm Sample with PF Marginalize out since they are conditionally independent given

},...,1{ Lt NL

LCt NiNiM ,...,1},,...,1{)( ))(( ttt LMfY

tL :1

)(iM t

tL :1