Kalman filtering for maximum likelihood estimation given corrupted observations.

Similar documents
An EM algorithm for maximum likelihood estimation given corrupted observations. E. E. Holmes, National Marine Fisheries Service

Data Fusion using Kalman Filter. Ioannis Rekleitis

Linear Gaussian State Space Models

References are appeared in the last slide. Last update: (1393/08/19)

Estimation of Poses with Particle Filters

3.1.3 INTRODUCTION TO DYNAMIC OPTIMIZATION: DISCRETE TIME PROBLEMS. A. The Hamiltonian and First-Order Conditions in a Finite Time Horizon

Tracking. Many slides adapted from Kristen Grauman, Deva Ramanan

GMM - Generalized Method of Moments

Probabilistic Robotics SLAM

Applications in Industry (Extended) Kalman Filter. Week Date Lecture Title

Introduction to Mobile Robotics

Georey E. Hinton. University oftoronto. Technical Report CRG-TR February 22, Abstract

Two Popular Bayesian Estimators: Particle and Kalman Filters. McGill COMP 765 Sept 14 th, 2017

Tracking. Many slides adapted from Kristen Grauman, Deva Ramanan

Modal identification of structures from roving input data by means of maximum likelihood estimation of the state space model

Probabilistic Robotics SLAM

USP. Surplus-Production Models

Introduction D P. r = constant discount rate, g = Gordon Model (1962): constant dividend growth rate.

ACE 564 Spring Lecture 7. Extensions of The Multiple Regression Model: Dummy Independent Variables. by Professor Scott H.

State-Space Models. Initialization, Estimation and Smoothing of the Kalman Filter

Testing for a Single Factor Model in the Multivariate State Space Framework

Simulation-Solving Dynamic Models ABE 5646 Week 2, Spring 2010

R t. C t P t. + u t. C t = αp t + βr t + v t. + β + w t

Probabilistic Robotics

Comparing Means: t-tests for One Sample & Two Related Samples

Almost Unbiased Estimator for Estimating Population Mean Using Known Value of Some Population Parameter(s)

Financial Econometrics Kalman Filter: some applications to Finance University of Evry - Master 2

Anno accademico 2006/2007. Davide Migliore

ALMOST UNBIASED RATIO AND PRODUCT TYPE EXPONENTIAL ESTIMATORS

KINEMATICS IN ONE DIMENSION

Lecture 33: November 29

Zürich. ETH Master Course: L Autonomous Mobile Robots Localization II

Recursive Least-Squares Fixed-Interval Smoother Using Covariance Information based on Innovation Approach in Linear Continuous Stochastic Systems

Speech and Language Processing

Hidden Markov Models

Testing the Random Walk Model. i.i.d. ( ) r

2.160 System Identification, Estimation, and Learning. Lecture Notes No. 8. March 6, 2006

L07. KALMAN FILTERING FOR NON-LINEAR SYSTEMS. NA568 Mobile Robotics: Methods & Algorithms

) were both constant and we brought them from under the integral.

Modeling Economic Time Series with Stochastic Linear Difference Equations

The Optimal Stopping Time for Selling an Asset When It Is Uncertain Whether the Price Process Is Increasing or Decreasing When the Horizon Is Infinite

Bias in Conditional and Unconditional Fixed Effects Logit Estimation: a Correction * Tom Coupé

A Bayesian Approach to Spectral Analysis

Block Diagram of a DCS in 411

Note: For all questions, answer (E) NOTA means none of the above answers is correct.

Using the Kalman filter Extended Kalman filter

Application of a Stochastic-Fuzzy Approach to Modeling Optimal Discrete Time Dynamical Systems by Using Large Scale Data Processing

Review - Quiz # 1. 1 g(y) dy = f(x) dx. y x. = u, so that y = xu and dy. dx (Sometimes you may want to use the substitution x y

The fundamental mass balance equation is ( 1 ) where: I = inputs P = production O = outputs L = losses A = accumulation

4.1 - Logarithms and Their Properties

Chapter 5. Heterocedastic Models. Introduction to time series (2008) 1

Solutions to the Exam Digital Communications I given on the 11th of June = 111 and g 2. c 2

A Dynamic Model of Economic Fluctuations

מקורות לחומר בשיעור ספר הלימוד: Forsyth & Ponce מאמרים שונים חומר באינטרנט! פרק פרק 18

Exponential model. The Gibbs sampler is described for ecological examples by Clark et al. (2003, 2004) and

3.1 More on model selection

5.1 - Logarithms and Their Properties

13.3 Term structure models

Notes on Kalman Filtering

Appendix to Creating Work Breaks From Available Idleness

Econ107 Applied Econometrics Topic 7: Multicollinearity (Studenmund, Chapter 8)

Vectorautoregressive Model and Cointegration Analysis. Time Series Analysis Dr. Sevtap Kestel 1

20. Applications of the Genetic-Drift Model

On Measuring Pro-Poor Growth. 1. On Various Ways of Measuring Pro-Poor Growth: A Short Review of the Literature

Anti-Disturbance Control for Multiple Disturbances

EXPONENTIAL PROBABILITY DISTRIBUTION

Excel-Based Solution Method For The Optimal Policy Of The Hadley And Whittin s Exact Model With Arma Demand

Ensamble methods: Bagging and Boosting

Chapter 2. First Order Scalar Equations

Diebold, Chapter 7. Francis X. Diebold, Elements of Forecasting, 4th Edition (Mason, Ohio: Cengage Learning, 2006). Chapter 7. Characterizing Cycles

Estimation Uncertainty

Article from. Predictive Analytics and Futurism. July 2016 Issue 13

Lab #2: Kinematics in 1-Dimension

Licenciatura de ADE y Licenciatura conjunta Derecho y ADE. Hoja de ejercicios 2 PARTE A

Chapter 4. Truncation Errors

ESTIMATION OF DYNAMIC PANEL DATA MODELS WHEN REGRESSION COEFFICIENTS AND INDIVIDUAL EFFECTS ARE TIME-VARYING

Lecture 20: Riccati Equations and Least Squares Feedback Control

Vehicle Arrival Models : Headway

Problem Set 5. Graduate Macro II, Spring 2017 The University of Notre Dame Professor Sims

You must fully interpret your results. There is a relationship doesn t cut it. Use the text and, especially, the SPSS Manual for guidance.

MA 366 Review - Test # 1

Lecture #8 Redfield theory of NMR relaxation

72 Calculus and Structures

Solutions from Chapter 9.1 and 9.2

PENALIZED LEAST SQUARES AND PENALIZED LIKELIHOOD

Sliding Mode Controller for Unstable Systems

in Engineering Prof. Dr. Michael Havbro Faber ETH Zurich, Switzerland Swiss Federal Institute of Technology

Role of Material Functions in Rheological Analysis. Role of Material Functions in Rheological Analysis

Age (x) nx lx. Age (x) nx lx dx qx

CS 4495 Computer Vision Tracking 1- Kalman,Gaussian

Linear Dynamic Models

Augmented Reality II - Kalman Filters - Gudrun Klinker May 25, 2004

Robot Motion Model EKF based Localization EKF SLAM Graph SLAM

Econ Autocorrelation. Sanjaya DeSilva

Uncertainty & Localization I

Properties of Autocorrelated Processes Economics 30331

Distribution of Estimates

Probabilistic Robotics Sebastian Thrun-- Stanford

Financial Econometrics Jeffrey R. Russell Midterm Winter 2009 SOLUTIONS

Transform Techniques. Moment Generating Function

Transcription:

alman filering maimum likelihood esimaion given corruped observaions... Holmes Naional Marine isheries Service Inroducion he alman filer is used o eend likelihood esimaion o cases wih hidden saes such as when observaions are corruped and he rue populaion size is unobserved. he following algorihm is based on secion 3.4 in Harve 989 which was used b Lindle 003 esimaion populaion processes. he alman filer is well-known and widel used in engineering and compuer science applicaions. here are a muliude of books on he alman filer including Harve 989. One of he more penerable inroducions of he alman filer alone bu no on maimum likelihood esimaion is chaper of Mabeck 979. he sae-space model he diffusion approimaion a sochasic eponenial growh model can be wrien as a linear sae space model wrien in he noaion familiar in he engineering lieraure: B w where w ~ normal0 Q [] v where v ~ f 0 R [] where log N is he rue log populaion size and log O is he log observaions of he populaion size. B is µ he mean populaion growh rae. Q is he σ oherwise known as he process error or environmenal variabili. R is he variabili associaed wih sampling error or oher non-process error. Onl is observed; he underling parameers B Q and R and he underling rue populaion size is hidden. If we make he assumpion ha v is normall disribued hen he model is a linear Gaussian sae-space model.

We can calculae he probabili of he observed ime series as follows: [3] where is he probabili of given all he observaions bee ime and 0. Denoe he epeced value of as ~. he condiional probabili is disribued normal wih a mean ~ and some variance denoed which depends on he paricular parameers Ψ B Q R ha generaed he daa. hus he probabili of he ime series given a paricular se of parameers Ψ is Ψ d / ~ ep ϑ π [4] from he probabili densi of a normal wih mean ~ and variance. he log likelihood of ψ given he daa is Ψ L ~ log log log π a consan. [5] or qn. 5 we need esimaes of ~ and. Observe from qn. ha R [6] he alman filer below gives opimal esimaes of and which are hen used in qn. 5 o calculae he log likelihood of ψ.

he maimum likelihood esimaes of B Q and R are found b using some pe of maimizaion rouine on qn. 6 o find he se of parameers ψ B Q R ha maimize he likelihood. Malab code his algorihm is given a he end of his appendi. he alman filer irs some noaion: ] [ ] [ he alman recursion: Sar a and sep ward o. Assume an iniial π and iniial 0 o sar he recursion. One could le hese be free variables and find he maimum likelihood values when maimizing qn. 6 bu ha is no done here and he algorihm should no be ver sensiive o hese saring values. A each sep compue: > > R Q B π his is he well-known alman filer bu i looks a lile differen han wha ou ll see in engineering es. irs generall i is assumed ha is a series of measuremens from muliple insrumens hus he alman filer is alwas wrien in mari m. Here since is one measuremen i can be wrien in scalar m. Second he alman filer is usuall presened

he model A Bu w filer is simplified quie a bi. C v. In his applicaion A C and u so he References Harve A. C. 99. orecasing srucural ime series models and he alman filer. Cambridge Universi ress Cambridge U. Mabeck. S. 979. Sochasic models esimaion and conrol. olume. Academic ress New York USA. Malab code funcion [musigmasigmanp]kalman_essdaa logdaa; %Sar wih some reasonable iniial parameer esimaes muesmean:end-:end-; mpvar:end-:end-; mp4var5:end-:end-4; sigmaesmp4-mp/3; sigmanpesvar:end-:end--ma0.000sigmaes/; pima0.000sigmaesma0.000sigmanpes; %var of %log ransm he variances so he search algorihm doesn give negaive % variances sarvals[mues logma0.000sigmaes logma0.000sigmanpes]; %fminsearch is a Nelder-Mead minimizaion malab funcion afminsearch'kalman_loglik'sarvals[]pi; MLmuesa; MLsigmaesepa; MLsigmanpesepa3; funcion negloglik kalman_loglikvini lengh; B v; %mu Q epv; %s R epv3; %snp %iniialize zeros; zeros; zeros; zeros; zeros; zeros; Jzeros; zeros;

zeros; vzeros; %ward pass ges ou [ given :] 0ini; 0; : if ini; %denoes _^0 ; %denoes _^0 else - B; % denoes _^-; Harve 3..a - Q; %Harve 3..b end /R; R; v -; * - ; %Harve 3..3a -*; %Harve 3.3.3b end %Calculae negaive log likelihood negloglik /*sumv.^./ /*sumlogabs /*log*pi;