Probabilistic Graphical Models

Transcription

1 Probabilistic Graphical Models Brown University CSCI 2950-P, Spring 2013 Prof. Erik Sudderth Lecture 12: Gaussian Belief Propagation, State Space Models and Kalman Filters Guest Kalman Filter Lecture by Jason Pacheco Some figures courtesy Michael Jordan s draft textbook, An Introduction to Probabilistic Graphical Models

2 Pairwise Markov Random Fields p(x) = 1 Z Y (s,t)2e st(x s,x t ) Y s2v s(x s ) x 1 x 2 Simple parameterization, but still expressive and widely used in practice Guaranteed Markov with respect to graph Any jointly Gaussian distribution can be represented by only pairwise potentials x 3 x 4 x 5 E V Z set of undirected edges (s,t) linking pairs of nodes set of N nodes or vertices, {1, 2,...,N} normalization constant (partition function)

3 Belief Propagation (Integral-Product) BELIEFS: Posterior marginals x t MESSAGES: Sufficient statistics m ts (x s ) / Z ˆp t (x t ) / t (x t ) Y (t) x t st(x s,x t ) t (x t ) u2 (t) m ut (x t ) neighborhood of node t (adjacent nodes) Y u2 (t)\s m ut (x t ) x s x t I) Message Product II) Message Propagation

4 Gaussian Distributions Simplest joint distribution that can capture arbitrary mean & covariance Justifications from central limit theorem and maximum entropy criterion Probability density above assumes covariance is positive definite ML parameter estimates are sample mean & sample covariance

5 Gaussian Conditionals & Marginals For any joint multivariate Gaussian distribution, all marginal distributions are Gaussians, and all conditional distributions are Gaussians

6 Partitioned Gaussian Distributions Marginals: Conditionals: Intuition: p(x 1 x 2 )= p(x 1,x 2 ) p(x 2 ) / exp 1 2 (x µ)t 1 (x µ)+ 1 2 (x 2 µ 2 ) T 1 22 (x 2 µ 2 )

7 Gaussian Conditionals & Marginals p(x 1 x 2 )=N x 1 µ (x 2 µ 2 ), 2 1(1 2 )

8 Gaussian Graphical Models x N (µ, ) J = 1

9 Gaussian Potentials

10 Interpreting GMRF Parameters var ( x s x N(s) ) =(Js,s ) 1 ρ st N(s,t) cov ( ) x s,x t x N(s,t) var ( ) ( ) = x s x N(s,t) var xt x N(s,t) J s,t Js,s J t,t

11 Gaussian Markov Properties x N (µ, ) J = 1 Theorem 2.2. Let x N (0,P) be a Gaussian stochastic process which is Markov with respect to an undirected graph G =(V, E). Assume that x is not Markov with respect to any G =(V, E ) such that E E, and partition J = P 1 into a V V grid according to the dimensions of the node variables. Then for any s, t V such that s t, J s,t = Jt,s T will be nonzero if and only if (s, t) E.

12 Inference with Gaussian Observations

13 Linear Gaussian Systems Marginal Likelihood: Posterior Distribution: Gaussian BP: Complexity linear, not cubic, in number of nodes

14 Belief Propagation (Integral-Product) BELIEFS: Posterior marginals x t MESSAGES: Sufficient statistics m ts (x s ) / Z ˆp t (x t ) / t (x t ) Y (t) x t st(x s,x t ) t (x t ) u2 (t) m ut (x t ) neighborhood of node t (adjacent nodes) Y u2 (t)\s m ut (x t ) x s x t I) Message Product II) Message Propagation

15 Gaussian Belief Propagation The natural, canonical, or information parameterization of a Gaussian distribution arises from quadratic form: 1 N (x #, ) / exp 2 xt x + # T x Gaussian BP N represents messages and marginals as: N # = 1 = 1 µ m ts (x s )=αn 1 (ϑ ts, Λ ts ) p (x s y) =N 1 (ϑ s, Λ s )

16 Gaussian Belief Propagation y 1 y 2\1 y 1 1 y 1 0 x 4 x 1 x 2 y 1 2 x 1 x 3 y 1 3 y 4\1 y 3\1 Gaussian BP represents messages and marginals as: N N m ts (x s )=αn 1 (ϑ ts, Λ ts ) p (x s y) =N 1 (ϑ s, Λ s ) Gaussian BP belief updates then have a simple form: ϑ n s = Cs T Rs 1 y s + p (x s ys n )=αp(y s x s ) m n ts (x s ) t N(s) p (y s x s )=αn 1( ) Cs T Rs 1 y s,cs T Rs 1 C s t N(s) Λ n s = Cs T Rs 1 C s + t N(s) ϑ n ts Λ n ts

17 Gaussian Belief Propagation y 1 1 y 1 0 y 1 3 y 1 2 x 1 x s x t ϑ = m n ts (x s )=α ψ s,t (x s,x t ) p (y t x t ) 0 x t ψ s,t (x s,x t ) p (y t x t ) Ct T Rt 1 y t + u N(t)\s ϑn 1 ut u N(t)\s Λ = m n 1 J s(t) J t,s ( ) m n 1 ut (x t ) dx t u N(t)\s ut (x t ) N 1( ϑ, Λ ) \ J t(s) + C T t R 1 t J s,t C t + u N(t)\s Λn 1 ut

18 Gaussian Belief Propagation y 1 1 y 1 0 y 1 2 x 1 y 1 3 x s x t M = A B C D (A BD M 1 = C) 1 (A BD 1 C) 1 BD 1 D 1 C (A BD 1 C) 1 D 1 + D 1 C (A BD 1 C) 1 BD 1 = A 1 + A 1 B (D CA 1 B) 1 CA 1 A 1 B (D CA 1 B) 1 (D CA 1 B) 1 CA 1 (D CA 1 B) 1 Matrix Inversion Lemma:

19 Gaussian Belief Propagation y 1 1 y 1 0 y 1 2 x 1 y 1 3 x s x t m n ts (x s )=α ϑ n ts Λ n ts = J s,t = J s(t) J s,t x t ψ s,t (x s,x t ) p (y t x t ) J t(s) + C T t R 1 t C t + u N(t)\s Λ n 1 ut J t(s) + C T t R 1 t C t + u N(t)\s N u N(t)\s 1 Λ n 1 ut m n 1 ut C T t R 1 t y t + 1 J t,s (x t ) dx t u N(t)\s ϑ n 1 ut

20 State Space Models Like HMM, but over continuous variables... Plant Equations:

21 State Space Models Dynamics: Process Noise Same as,

22 State Space Models Measurement: Measurement Noise Same as, This is known as the Linear Gaussian assumption.

23 Example: Target Tracking Target State Position Measurement Velocity Constant velocity dynamics (1st order diffeq) Lower-dimensional measurement

24 Example: Target Tracking

25 Correspondence With Factor Analysis Factor Analysis: State Space:

26 Inference for State Space Model Broken into 2 parts Filtering (Forward pass): Smoothing (Backward pass) Why filter? From signal processing Filters out system noise to produce an estimate

27 Conditional Moments Everything is Gaussian Can focus on mean / variance computations Conditional Mean Conditional Variance

28 Kalman Filter Two recursive updates: 1) Time update: best guess before seeing measurement) 2) Measurement Update: after measurement

29 Kalman Filter 1) Time update: 2) Measurement Update:

30 Kalman Filter: Time Update Recall that, so the conditional mean recursion is, Similar for covariance, Zero Noise In Expectation

31 Gaussian Conditionals For any joint multivariate Gaussian distribution, all conditional distributions are Gaussians

32 Gaussian Conditionals Gaussian joint distribution, Inverse Covariance Conditional is Gaussian with parameters, Conditional Covariance

33 Kalman Filter: Measurement Update Form the joint over as, Time Update Compute conditional, Measurement Equation

34 Kalman Filter: Measurement Update Quantity called the Kalman Gain Matrix Because it multiplies observation, i.e. produces gain Update takes linear combination of predicted mean and observation, weighted by predicted covariance

35 Kalman Filter Consider the covariance updates, Independent of observed measurements This is a property of Gaussians in general Only depend on process and measurement noise Can be computed offline

36 Kalman Filter

37 Information Filter Recall Gaussian has equivalent canonical parameterization Sometimes called Information form Inverse Covariance Canonical Mean Recursive updates follow definitions Matrix condition is reciprocal of condition of its inverse

38 Information Filter Define to, focus on precision update Matrix Inversion Lemma Measurement update Wait a minute...this looks familiar...

39 Gaussian Belief Propagation Message Update: Belief Update:

40 Gaussian BP => Kalman Filter Time Update: Measurement Update:

41 Smoother Combines forward and backward probabilities to produce full marginal posterior. Similar to inference on an HMM (forwardbackward algorithm)

42 Smoother Can we just invert the dynamics, and run Kalman filter backwards? x 1... x T-2 x T-1 x T y 1 y T-2 y T-1 y T No, w t is no longer independent of the past state (e.g. x t+1,, x T )

43 Unconditional Distribution Marginal distributions of a Gaussian are Gaussian Where, From zero-mean noise assumption Covariance computed recursively Does not depend on means

44 Smoother Given the unconditional marginal, Form the unconditional over Solve for reverse dynamics Run filter backwards with new dynamics

45 Smoother x 1... x T-2 x T-1 x T y 1 y T-2 y T-1 y T

46 Smoother Forward Conditional Backward Conditional Gaussian closed under multiplication Multiply to produce full smoothed marginal Kalman filter + smoother equivalent to Gaussian BP

47 Summary Kalman filter is optimal filter for Linear Gaussian State Space model Smoother provides full marginal inference Gaussian BP produces equivalent algorithm Correspondence clearly shown in Information form of the filter