1. Authors: Jarett Hailes, Mike Kouritzin, Jonathan Wiersma Hidden in the Noise: Tracking and Prediction in Complex Systems

2. Outline Particle Filtering Overview –Fish Tracking Example Selective Resampling Particle Filter –Multi-Target Tracking Problem Refining Grid Stochastic Filter –Performer Tracking & Prediction Problem Parameter Estimation in Random Environments Infinite Dimensional Exact Filter –Pod Tracking Problem –Financial Applications

3. Conditional Distribution )),...,(|( 1 tt YYXP  Signal: tttt dBXdtXdX)()(  Observation: ),( ttt VXhY  Filtering Overview Optimal Filter   ),...,(| 1tt YYXE 

4. Filtering Procedure 1.Modeling an unobserved signal. 2.Modeling the partial/noisy/distorted observation. 3.Filter outputs a conditional distribution estimate of the signal’s past/present/future state, based on observations. + = Signal & Observation Model Noisy Observations Filter’s Estimate Real Time

5. Sophisticated Monte-Carlo method that involves modeling signal/observation process Particles are copies of signals that evolve independently of one another between observations Weight particles based on state and observation based information Non Linear Particle Filters

6. N ∞ Weighted Particle Filters Optimal Filter           2 2 )( )( exp)(    YtYt hh W i t i t i t What’s Your Best Estimate? )),...,(|( 1tt YYXP       N i i t N i i t W W i t 1 1 )( )()(    

7.      N i N i (A) t tt W W YYAXP 1 1 1 )( δ )( )),...,(|( t    A i i t  i

8. Static Grid Filter Observation: Y t Observation: Y t+1 Size = Weight

9. Static Grid Filter Observation: Y t+1 Size = Weight

10. Adaptive Grid Filters Resampling

11. Adaptive Interacting Filter Time t - t i t YWby calculatedweight)(       N i tt i t N YYXP 1 1 )( 1 ),...,(|(   Time t 5 2 6 2 Resampling

12. ),...,| ( 111ttt N YYAXAXP   Resampling Issues N ∞ ),...,|()...,|()|( 12122111ttt YYAXPYYAXPYAXP     N i tt AA N i t i 1 ),...( )...( 1 1  

13. Historically Adaptive Grid Filters Mitacs-Pints Branching (MIBR) filter first historically adaptive filter (Fleischmann, Kouritzin 1999) Historical Adaptive Interacting Filter (Del Moral, Kouritzin, Miclo 2001) follows similar procedure as Adaptive Interacting Filter ),...,| ( 111 ttt N YYAXAXP      N i tt AA N i t i 1 ),...( )...( 1 1  

14. MIBR Filter Time Observation i

15. Fish Tracking Problem

16. Fish Model tttDttt dXXdvdt L XdX  )()() 2 (     L = L 1 x L 2 The MIBR method will track the fish

18. Selective Resampling Particle Filter (SERP Filter) Most recent particle filter developed at PINTS Dramatically outperforms any other particle filter Feedback control ρ determines extent of resampling ρ is determined via solution to stochastic control problem

19. 2 4 3 1) Resample Particles 2) Evolve Particles 3) Calculate Weights given new observation until W(ξ t i ) < ρ W(ξ t j ) all i, j )()( )( j t i t j t WW W    Prob: 22 )()( )( j t i t i t WW W    )( j t W  )( i t W  2 )()( i t j t WW 

20. Selective Resampling Particle Filter (SERP Filter) Naturally efficient data structures Less resampling noise Optimal amount of resampling through feedback mechanism

21. Multiple Target Tracking Goal: Determine the number of targets present in a given area and track them effectively Need to be able to model potential interactions between targets

22. Multiple Ships Lost at Sea Sample problem: An unknown number of dinghies in bounded space need to be identified and tracked

23. Multiple Ships Lost at Sea Signal Model : Ship is a nonlinear system with six state components: (x,y) position, orientation, and their respective speeds Have a maximum/minimum speed Speeds affected by type of motion: Rowing,Drifting,Motoring – Seventh state Orientation drifts towards 0 degrees

24. Weakly Interacting Targets Attraction-Repulsion Field y             motorized rowing, adrift, :state chi norientatioin change positionin change positionin change norientatio position   ... x y x X t

25. Attraction Repulsion Field

26. Particle Content Particles contain many targets within them Can be resampled with particles that contain other targets

27. Ships Lost at Sea Observation Model : We only have noisy discrete-time observations from a high-altitude sensor on a helicopter 11 10 11 1 0 0 0 00 0 0 0 0 00 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00 0 l m   ml tt mlml t VXhY,,, )(

30. Refining Grid Stochastic Filter (REST) Filter - Given a signal that evolves on regular Euclidean subset - Divide signal state space into a finite number of cells In general N 1 x N 2 x … x N d cells N1N1 N2N2

31. Each cell contains: 42 - Particle count - Associated Rate

32. Refining Grid Stochastic Filter (REST Filter) Particles used to approximate unnormalized conditional distribution

33. 1 2 -2 1 Cell Rates Cell rates are used to calculate net birth (death rate) in a cell Rates are determined by cell’s particle count and immediate neighbour’s rates = Net Birth Rate +1

34. Net birth rates are used to mimic signal movement and incorporate observation information Net Birth Rates OBSERVATIONOBSERVATION +=

35. N2N2 N1N1 Dynamic Cell Sizing Zoom in: N 1 Zoom out: N 1

36. Dynamic Cell Sizing Example

37. Tree Node Cell Node 4 2 1 6 17 3 Observation: -2 Particles 14

38. 4 13 1 2 12 4 2 2 Splitting Cells… Tree Node Cell Node

39. REST Advantages Less simulation noise than particle filters Dynamic cell sizing, inherent parameter estimation Dynamic domain problems

40. Performer Problem - Acoustic tracking system designed to have lighting equipment follow performer on large stage - Due to mechanical lags, system must be able to predict performer’s future position based on current state

41. Audience

42. Performer Model θ Audience

43. Observation Model      null    otherwise )1,0( if ),( )()+( (),( 222 pUWSXh Y zSySxSSXh mt l t t l z l y l x l t mm m m  S 1 (x,y,z) S 3 (x,y,z) S 4 (x,y,z) S 2 (x,y,z)

45. Parameter Estimation for Random Environments tttt dBXdtXWdX)()( 

46. Diffusions with Unknown Parameters –Combined state and parameter estimation –Filter performance –Example Problem Diffusions in Random Environments –Motivation –Nonparametric estimation method –Particle filters and simulations

47. Common method for filtering problem Filtering Problem Common Parameter Estimation Method Within Filtering: –Conditional likelihood maximization or EM algorithm –Expensive computation

48.    0 )( n i (i+1) a i x i xf Example Problem Recursive Algorithm Generalized Method of Moments (GMM) Estimate parameters via: )1,0()( 2ktk NXhY k 

49. Invariant measure method Idea: –Find moments of invariant measure by ergodic theorem - GMM –Obtain unknown parameters through solving non-singular finite-order linear equations Invariant Measure:

50. Invariant measure method (Cont) Ergodic Theorem: Expression of Moments of

51. Simulation result: parameter estimation

52. Simulation result: filtering via branching particle method

53.            t i iiit YYXhEY t 0 2 1111 )],...,(|)([ 1 1 minarg ˆ   Recursive Algorithm ) ˆ ( ˆˆ 111   tttttt YYAP 

54. Example 01 1 Dimensional Signal t d dt dBxadxaxaadX    ))1(...32( 2 1 2 210 )1( }1{ 2 1 )0( }0{ 2 1 )()( tttt dXdX 

56. Particle filters Initialization Particles are independently initialized Evolution: Particles are evolved according to approximate SDE with current parameter estimate Observation Arrival Parameter Estimation Calculate moments via solving linear equations or recursive method Selection: Particles are branched

57. Part II: Diffusions in random environments Motivation: Tracking problem of a dinghy lost on lake –In calm lakes dinghy moves as reflecting diffusion process –Lake swells provide random drift Signal Model:

58. Noisy observation and filtering problem Observation Sequence: random vector fields Filtering Problem: Let. For a continuous function, the filtering problem is to evaluate

59. Nonparametric estimation method Idea: –Perform functional estimations for the fixed unknown path of the random medium –Use obtained estimators to construct approximate filters Ergodic Theorem:

60. Approximate diffusions By trigonometric Fourier series approximation we get smooth estimators for. Define approximate diffusions Approximate filters

61. Filter Robustness Bhatt & Karandikar (1999) robustness yields Budhiraja & Kushner (2002) long time behavior gives as n,T ∞ in any way

62. Particle filters (branching) Initialization: Particles are independently initialized Evolution: Particles are evolved according to approximate SDE with current parameter estimates Observation Arrival Functional Estimation: Calculate functional environment estimate Selection: Particles are branched

63. Simulation Results:

64. Conclusions New methods of filtering for reflecting diffusions in random medium or with unknown parameters Have no access to true random medium or parameters, but contend with noisy estimates Computer workable methods Simulations experimentally validate our methods

65. Further investigations Generalize nonparametric estimation method to handle signals in random scenery and fractional BM Modify parameter estimation procedure to get estimates at all frequencies Examine performance of particle filters on more complicated signal models

66. Infinite Dimensional Exact Filter (IDEX) Kalman, Benes, and Daum exact filters propagate finite dimensional sufficient statistic Finite dimensionality limits possible observation models Computer efficient convolution allows for more general infinite dimensional exact filters Developed using explicit solutions, Feynman-Kac formulas, and Ricatti equations Certain nonlinear drift, dispersion coefficients possible

67. IDEX Formula a) b) c) d) e) )())()(|)( 1 xFxYP(Xx jt j   )( ˆ  FFT               2 exp)( ˆ )exp( 1 )( ˆ 2   aa b i a ) ˆ (  IFFT ),()())(| 1 xtHvxYP(X jxjt     where             xdzzmztxtH x jj    )()(exp),( 

68. IDEX Advantages - No simulation noise - General observation assumption

69. Pod Tracking Problem Pod Sensor Second Sensor

70. Stratonovich SDE ssss dWdWXdsXhdX  )()(             3 2 1 x x x X s 233 33 : : x h    Process WienerD2:  s W

71. Vector Fields Constraining Signal to Sphere    2 0 0 0 )()()( k t k sskt dWXfAXfXf  )()())(( 3 1 0 xfxhxfA i ii    1,2j )()())(( 3 1    xfxxfA i iijj  0  fA k 2 3 2 2 2 1 )(xxxXf 

72. sensor pod ofposition restingr r tonormal plane of basis},{ )(proj 21 },{ 21      ttt WXY Pod Observation Model

73. θ φ

74. Pod Tracking Developments - Workable explicit solution - Eliminate approximation errors - More realistic manifolds - Cantilever equations

76. Financial Applications Problem: Determine model parameters for option pricing Idea: Model Signal = function of volatility – combined tracking and parameter estimation. The signal’s formula is: with parameter  = time constant of signal process         W eZ e Z s ts t t d 00           2 1,0Νassuming Z 0

77. Signal is unobservable volatility in stock price of publicly traded company Observations also parameterized with: –  = measure of mean volatility in log scale –  = amplitude of volatility noise –  = index of stable law for noise –m = drift

78. Stock Prices: P 0,...,P n Volatility P[g(Z t )|Y k, t k ≤ t] Y k = log(P k /P k-1 ) Volatility Observation Model     parameter withv. r. stable symmetricstandard zg kkwhere t gm S e t S ZY k z k k k k           ...,3,2,1,