Particle Gibbs with Ancestor Sampling Fredrik Lindsten ⋆ , Michael I. Jordan † , Thomas B. Schön ‡ Chamonix, January 6, 2014 ⋆ Division of Automatic Control Linköping University, Sweden † Departments of EECS and Statistics University of California, Berkeley, USA ‡ Department of Information Technology Uppsala University, Sweden AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
Identification of state-space models 2(24) Consider a nonlinear discrete-time state-space model, x t ∼ f θ ( x t | x t − 1 ) , y t ∼ g θ ( y t | x t ) , and x 1 ∼ π ( x 1 ) . We observe y 1: T = ( y 1 , . . . , y T ) and wish to estimate θ . • Frequentists: Find ˆ θ ML = arg max θ p θ ( y 1: T ) . - Use e.g. the Monte Carlo EM algorithm. • Bayesians: Find p ( θ | y 1: T ) . - Use e.g. Gibbs sampling. AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
Gibbs sampler for SSMs 3(24) Aim: Find p ( θ , x 1: T | y 1: T ) . MCMC: Gibbs sampling for state-space models. Iterate, • Draw θ [ k ] ∼ p ( θ | x 1: T [ k − 1 ] , y 1: T ) ; OK! • Draw x 1: T [ k ] ∼ p θ [ k ] ( x 1: T | y 1: T ) . Hard! Problem: p θ ( x 1: T | y 1: T ) not available! Idea: Approximate p θ ( x 1: T | y 1: T ) using a particle filter. AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
The particle filter 4(24) Weighting Resampling Propagation Weighting Resampling t − 1 ) = w j • Resampling: P ( a i t = j | F N t − 1 / ∑ l w l t − 1 . t ( x t | x a i 1: t = { x a i • Propagation: x i t ∼ R θ 1: t − 1 ) and x i 1: t − 1 , x i t } . t t • Weighting: w i t = W θ t ( x i 1: t ) . ⇒ { x i 1: t , w i t } N i = 1 AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
The particle filter 5(24) Algorithm Particle filter (PF) 1. Initialize ( t = 1 ): (a) Draw x i 1 ∼ R θ 1 ( x 1 ) for i = 1, . . . , N . (b) Set w i 1 = W θ 1 ( x i 1 ) for i = 1, . . . , N . 2. for t = 2, . . . , T : t ∼ Discrete ( { w j (a) Draw a i t − 1 } N j = 1 ) for i = 1, . . . , N . t ( x t | x a i (b) Draw x i t ∼ R θ 1: t − 1 ) for i = 1, . . . , N . t 1: t = { x a i (c) Set x i 1: t − 1 , x i t } and w i t = W θ t ( x i t 1: t ) for i = 1, . . . , N . AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
The particle filter 6(24) 1 0 −1 State −2 −3 −4 5 10 15 20 25 Time AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
Sampling based on the PF 7(24) approx. 1: T = x i 1: T | F N T ) ∝ w i ∼ p θ ( x 1: T | y 1: T ) . With P ( x ⋆ T we get x ⋆ 1: T 1 0.5 0 −0.5 −1 State −1.5 −2 −2.5 −3 −3.5 −4 5 10 15 20 25 Time AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
Conditional particle filter with ancestor sampling 8(24) Problems with this approach, • Based on a PF ⇒ approximate sample. • Does not leave p ( θ , x 1: T | y 1: T ) invariant! • Relies on large N to be successful. • A lot of wasted computations. Conditional particle filter with ancestor sampling (CPF-AS) Let x ′ 1: T = ( x ′ 1 , . . . , x ′ T ) be a fixed reference trajectory . • At each time t , sample only N − 1 particles in the standard way. • Set the N th particle deterministically: x N t = x ′ t . • Generate an artificial history for x N t by ancestor sampling. AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
Conditional particle filter with ancestor sampling 9(24) Algorithm CPF-AS, conditioned on x ′ 1: T 1. Initialize ( t = 1 ): (a) Draw x i 1 ∼ R θ 1 ( x 1 ) for i = 1, . . . , N − 1 . (b) Set x N 1 = x ′ 1 . (c) Set w i 1 = W θ 1 ( x i 1 ) for i = 1, . . . , N . 2. for t = 2, . . . , T : t ∼ Discrete ( { w j (a) Draw a i t − 1 } N j = 1 ) for i = 1, . . . , N − 1 . t ( x t | x a i (b) Draw x i t ∼ R θ 1: t − 1 ) for i = 1, . . . , N − 1 . t (c) Set x N t = x ′ t . (d) Draw a N t with P ( a N t = i | F N t − 1 ) ∝ w i t | x i t − 1 f θ ( x ′ t − 1 ) . 1: t = { x a i (e) Set x i 1: t − 1 , x i t } and w i t = W θ t ( x i t 1: t ) for i = 1, . . . , N . AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
The PGAS Markov kernel (I/II) 10(24) Consider the procedure: 1. Run CPF-AS ( N , x ′ 1: T ) targeting p θ ( x 1: T | y 1: T ) , 1: T = x i 1: T | F N T ) ∝ w i 2. Sample x ⋆ 1: T with P ( x ⋆ T . 3 2 1 State 0 −1 −2 −3 5 10 15 20 25 30 35 40 45 50 Time AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
The PGAS Markov kernel (II/II) 11(24) This procedure: • Maps x ′ 1: T stochastically into x ⋆ 1: T . • Implicitly defines a Markov kernel ( P N θ ) on ( X T , X T ) , referred to as the PGAS (Particle Gibbs with ancestor sampling) kernel. Theorem For any number of particles N ≥ 1 and for any θ ∈ Θ , the PGAS kernel P N θ leaves p θ ( x 1: T | y 1: T ) invariant, � θ ( x ′ 1: T ) p θ ( dx ′ p θ ( dx ⋆ P N 1: T , dx ⋆ 1: T | y 1: T ) = 1: T | y 1: T ) . F. Lindsten, M. I. Jordan and T. B. Schön , P . Bartlett, F . C. N. Pereira, C. J. C. Burges, L. Bottou and K. Q. Weinberger (Eds.), Ancestor Sampling for Particle Gibbs Advances in Neural Information Processing Systems (NIPS) 25 , 2600-2608, 2012. F. Lindsten, M. I. Jordan and T. B. Schön , Particle Gibbs with Ancestor sampling, arXiv:1401.0604 , 2014. AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
Ergodicity 12(24) Theorem Assume that there exist constants ε > 0 and κ < ∞ such that, for any θ ∈ Θ , t ∈ { 1, . . . , T } and x 1: t ∈ X t , W θ t ( x 1: t ) ≤ κ and p θ ( y 1: T ) ≥ ε . a Then, for any N ≥ 2 the PGAS kernel P N θ is uniformly ergodic. That is, there exist constants R < ∞ and ρ ∈ [ 0, 1 ) such that θ ) k ( x ′ ∀ x ′ � ( P N 1: T , · ) − p θ ( · | y 1: T ) � TV ≤ R ρ k , 1: T ∈ X T . a N.B. These conditions are simple, but unnecessarily strong; see (Lindsten, Douc, and Moulines. 2014). F. Lindsten, M. I. Jordan and T. B. Schön , Particle Gibbs with Ancestor sampling, arXiv:1401.0604 , 2014. AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
PGAS for Bayesian identification 13(24) Bayesian identification: PGAS + Gibbs Algorithm PGAS for Bayesian identification 1. Initialize: Set { θ [ 0 ] , x 1: T [ 0 ] } arbitrarily. 2. For k ≥ 1 , iterate: (a) Draw x 1: T [ k ] ∼ P N θ [ k − 1 ] ( x 1: T [ k − 1 ] , · ) . (b) Draw θ [ k ] ∼ p ( θ | x 1: T [ k ] , y 1: T ) . For any number of particles N ≥ 2 , the Markov chain { θ [ k ] , x 1: T [ k ] } k ≥ 1 has limiting distribution p ( θ , x 1: T | y 1: T ) . AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
ex) Stochastic volatility model 14(24) Stochastic volatility model, x t + 1 = 0.9 x t + v t , v t ∼ N ( 0, θ ) , e t ∼ N ( 0, 1 ) . y t = e t exp ( 1 2 x t ) , Consider the ACF of θ [ k ] − E [ θ | y 1: T ] . PG-AS, T = 1000 PG, T = 1000 1 N=5 1 N=5 N=20 N=20 N=100 N=100 0.8 0.8 N=1000 N=1000 0.6 0.6 ACF ACF 0.4 0.4 0.2 0.2 0 0 0 50 100 150 200 0 50 100 150 200 Lag Lag AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
PGAS vs. PG 15(24) 3 2 1 State 0 −1 −2 −3 5 10 15 20 25 30 35 40 45 50 Time PGAS PG 3 3 2 2 1 1 State State 0 0 −1 −1 −2 −2 −3 −3 5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50 Time Time AUTOMATIC CONTROL Fredrik Lindsten, Michael I. Jordan, Thomas B. Schön REGLERTEKNIK LINKÖPINGS UNIVERSITET Inference in nonlinear state-space models using Particle Gibbs with Ancestor Sampling
Recommend
More recommend