The Unscented Particle Filter

Transcription

1 The Unscented Particle Filter Rudolph van der Merwe (OGI) Nando de Freitas (UC Bereley) Arnaud Doucet (Cambridge University) Eric Wan (OGI)

2 Outline Optimal Estimation & Filtering Optimal Recursive Bayesian Solution Practical Solutions Gaussian approximations (EKF, UKF) Sequential Monte Carlo methods (Particle Filters) The Unscented Particle Filter The Unscented Transformation and UKF Applications of UT/UKF to Particle Filters Experimental Results Conclusions

3 Filtering General problem statement y y 2 1 y Observed p( y x ) Unobserved x 2 x 1 p( x x ) 1 x Filtering is the problem of sequentially estimating the states (parameters or hidden variables) of a system as a set of observations become available on-line.

4 Filtering Solution of sequential estimation problem given by Posterior density : p( X Y ) X = { x, x,, x } 1 2 Y = { y, y,, y } 1 2 By recursively computing a marginal of the posterior, the filtering density, p( x Y ) one need not eep trac of the complete history of the states.

5 Filtering Given the filtering density, a number of estimates of the system state can be calculated: Mean (optimal MMSE estimate of state) xˆ E x Y x p( x Y ) dx [ ] = = Mode Median Confidence intervals Kurtosis, etc.

6 State Space Formulation of System General discrete-time nonlinear, non-gaussian dynamic system state x = f( x, u, v ) y = hx (, u, n ) process noise nown input noisy observation measurement noise Assumptions : 1) States follow a first order Marov process p( x x, x,, x ) = p( x x ) ) Observations independent given the states p( y x, A) = p( y x )

7 Recursive Bayesian Estimation Given this state space model, how do we recursively estimate the filtering density? p( x Y ) = = = = = p( Y x) p( x) p( Y ) p( y, Y 1 x) p( x) p( y, Y ) 1 p( y Y 1, x) p( Y 1 x) p( x) p( y Y ) p( Y ) 1 1 p( y Y 1, x) p( x Y 1) p( Y 1) p( x ) p( y Y ) p( Y ) p( x ) p( y 1 1 x) p( x Y 1) p( y Y ) 1

8 Recursive Bayesian Estimation lielihood prior posterior p( x Y ) = p( y x) p( x Y 1) p( y Y ) 1 evidence Prior : p( x Y ) = p( x x ) p( x Y ) dx transition density given by process model (Propagation of past state into future before new observation is made.) Lielihood : defined in terms of observation model Evidence : p( y Y ) = p( y x ) p( x Y ) dx 1 1

9 Practical Solutions Gaussian Approximations Perfect Monte Carlo Simulation Sequential Monte Carlo Methods : Particle Filters Bayesian Importance Sampling Sampling-importance resampling (SIR)

10 Gaussian Approximations Most common approach. Assume all RV statistics are Gaussian. Optimal recursive MMSE estimate is then given by xˆ = ( prediction of x ) + κ [( observation of y ) ( prediction of y )] Different implementations : Extended Kalman Filter (EKF) : optimal quantities approximated via first order Taylor series expansion (linearization) of process and measurement models. Unscented Kalman Filter (UKF) : optimal quantities calculated using the Unscented Transformation (accurate to second order for any nonlinearity). Drastic improvement over EKF [Wan, van der Merwe, Nelson 2000]. Problem : Gaussian approximation breas down for most nonlinear real-world applications (multi-modal distributions, non-gaussian noise sources, etc.)

11 Perfect Monte Carlo Simulation Allow for a complete representation of the posterior distribution. Map intractable integrals of optimal Bayesian solution to tractable discrete sums of weighted samples drawn from the posterior distribution. N 1 = N δ i= 1 ( () i ) pˆ ( x Y ) x x x p( x Y ) () i IID... So, any estimate of the form [ ] E f( x ) = f( x ) p( x Y ) dx may be approximated by : [ ] 1 ( () i ( x ) ) N f x E f N i= 1

12 Particle Filters Bayesian Importance Sampling It is often impossible to sample directly from the true posterior density. However, we can rather sample from a nown, easy-tosample, proposal distribution, q( x Y ) and mae use of the following substitution p( x Y ) [ ( )] = ( ) ( ) q( x Y ) E f x f x dx w x p( x ) ( x Y ) p( Y ) q q x Y = = ( x ) = ( Y x ) p( x ) ( Y ) q( x Y ) ( ) f ( x ) q x Y dx ( x) ( ) q Y ( ) p p w p f ( x ) x Y dx

13 Particle Filters 1 [ ( )] = ( ) ( ) ( ) ( ) E f x f x w x q x Y dx p Y = = = ( ) f( x ) w ( x ) q x Y dx p( Y x ) p( x ) q ( x Y) ( x Y ) q dx q ( ) ( ) ( ) w ( x ) q( x Y ) dx ( ) [ ( ) ( )] x x x Y Eq( x Y ) [ w( x) ] f x w x q x Y dx E w f

14 Particle Filters q( x Y ) So, by drawing samples from, we can approximate expectations of interest by the following: [ ( x )] E f 1 N () i () i N w ( ) ( ) 1 x f x i= 1 N () i N w ( ) i= 1 x N i= 1 w () i () i ( x ) f ( x ) Where the normalized importance weights are given by w () i () i w( x ) ( x ) = N ( j) w j= 1 x ( )

15 Particle Filters Using the state space assumptions (1 st order Marov / observational independence given state), the importance weights can be estimated recursively by [proof in De Freitas (2000)] w = w 1 p y x p x x ( ) ( ) q 1 x X, Y ( ) 1 Problem with SIS is that the variance of the importance weights increase stochastically over time [Kong et al. (1994), Doucet et al. (1999)] To solve this, we need to resample the particles eep / multiply particles with high importance weights discard particles with low importance weights Sampling-importance Resampling (SIR)

16 Particle Filters Sampling-importance Resampling Maps the N unequally weighted particles into a new set of N equally weighted samples. { () ()} { ( ) 1 x i, i j, } w x N Method proposed by Gordon, Salmond & Smith (1993) and proven mathematically by Gordon (1994). cdf 1 N 1 ( j ) w t j resampled index p(i)

17 Particle Filters

18 Particle Filters Choice of Proposal Distribution w = w 1 p ( y ) ( ) x p x x 1 q( x X, Y ) 1 critical design issue for successful particle filter samples/particles are drawn from this distribution used to evaluate importance weights Requirements 1) Support of proposal distribution must include support of true posterior distribution, i.e. heavy-tailed distributions are preferable. 2) Must include most recent observations.

19 Particle Filters Most popular choice of proposal distribution does not satisfy these requirements though: q x X, Y = p x x ( ) ( ) 1 1 [Isard and Blae 96, Kitagawa 96, Gordon et al. 93, Beadle and Djuric 97, Avitzour 95] Easy to implement : w = w 1 p y x p x x ( ) ( ) 1 p x ( ) x 1 = w p ( y x ) 1 Does not incorporate most recent observation though!

20 Improving Particle Filters Incorporate New Observations into Proposal Use Gaussian approximation (i.e. Kalman filter) to generate proposal by combining new observation with prior q x X, Y = p x X, Y ( ) ( ) 1 G 1 = N ˆ,cov[ ] ( x x )

21 Improving Particle Filters Extented Kalman Filter Proposal Generation De Freitas (1998), Doucet (1998), Pitt & Shephard (1999). Greatly improved performance compared to standard particle filter in problems with very accurate measurements, i.e. lielihood very peaed in comparison to prior. In highly nonlinear problems, the EKF tends to be very inaccurate and underestimates the true covariance of the state. This violates the distribution support requirement for the proposal distribution and can lead to poor performance and filter divergence. We propose the use of the Unscented Kalman Filter for proposal generation to address these problems!

22 Improving Particle Filters Unscented Kalman Filter Proposal Generation UKF is a recursive MMSE estimator based on the Unscented Transformation (UT). UT : Method for calculating the statistics of a RV that undergoes a nonlinear transformation (Julier and Uhlmann 1997) UT/UKF : - accurate to 3 rd order for Gaussians - higher order errors scaled by choice of transform parameters. More accurate estimates than EKF (Wan, van der Merwe, Nelson 2000) Have some control over higher order moments, i.e. urtosis, etc. heavy tailed distributions!

23 Unscented Transformation b i Weighted sample mean x y a + - f b g y i Weighted sample covariance P x P y b i l q = + - c x x a P x a P i x x 8

24 The Unscented Transformation Actual (sampling) UT Linearized (EKF) covariance sigma points mean y= f( x ) ψ = f ( χ ) i i T y = f ( x) Py = A P A x true mean UT mean true covariance UT covariance transformed sigma points A T P x A f ( x) 9

25 Unscented Particle Filter Particle Filter UKF Proposal + Unscented Particle Filter

26 Experimental Results Synthetic Experiment Time-series process model : x x v 1 1 sin ( ωπ ) + = + + φ + process noise (Gamma) nonstationary observation model : y 2 φ x + n = φ x 2 + n > measurement noise (Gaussian)

27 Experimental Results Synthetic Experiment : (100 independent runs) Filter Extended Kalman Filter (EKF) MSE mean variance Unscented Kalman Filter (UKF) Particle Filter : generic Particle Filter : EKF proposal Unscented Particle Filter

28 Experimental Results Synthetic Experiment

29 Experimental Results Pricing Financial Options Options : financial derivative that gives the holder the right (but not obligation) to do something in the future. Call option : - allow holder to buy an underlying cash product - at a specified future date ( maturity time ) - for a predetermined price ( strie price ) Put option : - allow holder to sell an underlying cash product Blac Scholes partial differential equation Main industry standard for pricing options ris-free interest rate 2 f f f rs + 2 σ S = rf 2 t S S volatility of cash product option value value of underlying cash product

30 Experimental Results Pricing Financial Options Blac & Scholes (1973) derived the following pricing solution: C = SN ( d ) Xe N ( d ) c rtm 1 c 2 rtm P= SN ( d1) + Xe N ( d ) c c 2 N c d 1 = ln( / ) + ( + 2σ ) d = d σ t S X r t σ m t m (.) = cumulative normal distribution m

31 Experimental Results Pricing Financial Options State-space representation to model system for particle filters Hidden states : Output observations: Known control signals: r C t m, σ, P, S Estimate call and put prices over a 204 day period on the FTSE-100 index. Performance : normalized square error for one-step-ahead predictions NSE = y ˆ ( y )

32 Experimental Results Options Pricing Experiment : (100 independent runs) Option Type Call Put Algorithm Trivial Extended Kalman Filter (EKF) Unscented Kalman Filter (UKF) Particle Filter : generic Particle Filter : EKF proposal Unscented Particle Filter Trivial Extended Kalman Filter (EKF) Unscented Kalman Filter (UKF) Particle Filter : generic Particle Filter : EKF proposal Unscented Particle Filter NSE mean var

33 Experimental Results Options Pricing Experiment : UPF one-step-ahead predictions

34 Experimental Results Options Pricing Experiment : Estimated interest rate and volatility

35 Experimental Results Options Pricing Experiment : Probability distributions of implied interest rate and volatility

36 Particle Filter Demos Visual Dynamics Group, Oxford. (Andrew Blae) Tracing agile motion Tracing motion against camouflage Mixed state tracing

37 Conclusions Particle filters allow for a practical but complete representation of posterior probability distribution. Applicable to general nonlinear, non-gaussian estimation problems where standard Gaussian approximations fail. Particle filters rely on importance sampling, so the proper choice of proposal distribution is very important: Standard approach (i.e. transition prior proposal) fails when lielihood of new data is very peaed (accurate sensors) or for heavy-tailed noise sources. EKF proposal : Incorporates new observations, but can diverge due to inaccurate and inconsistent state estimates. Unscented Particle Filter : UKF proposal More consistent and accurate state estimates. Larger support overlap, can have heavier tailed distributions. Theory predicts and experiments prove significantly better performance.

38 The End