Upcoming SlideShare
×

# Dealing with intractability: Recent Bayesian Monte Carlo methods for dealing with intractable likelihoods

2,392 views
2,333 views

Published on

talk by Nicolas Chopin at CREST Statistics Seminar, 16/01/2011.
This is partly a review, partly a talk on recent research such as

http://arxiv.org/abs/1101.1528

Published in: Technology, Lifestyle
1 Like
Statistics
Notes
• Full Name
Comment goes here.

Are you sure you want to Yes No
• Be the first to comment

Views
Total views
2,392
On SlideShare
0
From Embeds
0
Number of Embeds
2
Actions
Shares
0
23
0
Likes
1
Embeds 0
No embeds

No notes for slide

### Dealing with intractability: Recent Bayesian Monte Carlo methods for dealing with intractable likelihoods

1. 1. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2 Dealing with intractability: recent advances in Bayesian Monte Carlo methods for intractable likelihoods N. CHOPIN1 CREST-ENSAE 1 joint work with S. BARTHELME, P.E. JACOB, & O.PAPASPILIOPOULOS N. CHOPIN Intractability 1/ 54
2. 2. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Outline 1 Background 2 ABC methods for generative models 3 MC2 type methods 4 State-Space models, PMCMC 5 SMC2 N. CHOPIN Intractability 2/ 54
3. 3. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Tractable models For a prototypic Bayesian model, deﬁned by (a) prior p(θ), and (b) likelihood p(y |θ), a standard approach is to sample from the posterior p(θ|y ) ∝ p(θ)p(y |θ). using the Metropolis-Hastings algorithm: Metropolis-Hastings From current point θn 1 Sample θp ∼ T (θn , dθp ) 2 With probability 1 ∧ r , take θn+1 = θp , otherwise θn+1 = θn , where p(θp )p(y |θp )T (θp , θn ) r= p(θn )p(y |θn )T (θn , θp ) This generates a Markov chain which leaves the posterior invariant. N. CHOPIN Intractability 3/ 54
4. 4. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Intractable models This generic approach cannot be applied in the following situations: 1 The likelihood reads p(y |θ) = C (θ)hθ (y ), where C (θ) is an intractable normalising constant; e.g. log-linear models, Ising models. 2 The likelihood p(y |θ) is an intractable integral p(y |θ) = p(y , x|θ) dx X of a tractable integrand; e.g. state-space models. 3 The likelihood is even more complicate, because it corresponds to some generative process (scientiﬁc models). Solutions to these problems involve auxiliary variables. N. CHOPIN Intractability 4/ 54
5. 5. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Outline 1 Background 2 ABC methods for generative models 3 MC2 type methods 4 State-Space models, PMCMC 5 SMC2 N. CHOPIN Intractability 5/ 54
6. 6. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Example of a generative model: reaction times Subject must choose between k alternatives. Evidence ej (t) in favour of choice j follows a Brownian motion with drift: τ dej (t) = mj dt + dWtj . Decision is taken when one evidence “wins the race”; see plot. Threshold for B Threshold for A Evidence for B Evidence for A 0 50 100 150 time (ms) N. CHOPIN Intractability 6/ 54
7. 7. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2ABC methods for generative models ABC stands for “Approximate Bayesian Computation”. In such algorithms, the auxiliary variable is an artiﬁcial dataset y ∼ p(y |θ). Denote the actual dataset y . Consider the simple rejection algorithm: Basic ABC Repeat 1 Sample θ ∼ p(θ). 2 Sample y ∼ p(y |θ). 3 Accept with probability Kε ( s(y ) − s(y ) ). where Kε (x) = K (x/ε), K is a kernel function, and s is a vector of “summary statistics”. N. CHOPIN Intractability 7/ 54
8. 8. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2ABC target This algorithm samples from: πε (θ, y ) ∝ p(θ)p(y |θ)Kε ( s(y ) − s(y ) ). and the marginal πε (θ) → p(θ|s(y )) as ε → 0. If s is suﬃcient, then the limit is the true posterior p(θ|s(y )) = p(θ|y ), but this is rarely possible unfortunately. N. CHOPIN Intractability 8/ 54
9. 9. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2MCMC-ABC One can instead derive a MCMC algorithm that sample from the same distribution. MCMC-ABC From current point (θn , yn ) 1 Sample θp ∼ T (θn , dθp ). 2 Sample y p ∼ p(y |θp ). 3 With probability 1 ∧ r , take (θn+1 , yn+1 ) = (θp , y p ), otherwise (θn+1 , yn+1 ) = (θn , yn ), where p(θp )Kε ( s(y p ) − s(y ) )T (θp , θn ) r= p(θn )Kε ( s(yn ) − s(y ) )T (θn , θp ) N. CHOPIN Intractability 9/ 54
10. 10. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Remarks on the KDE interpretation of ABC Having sampled N pairs (θi , y i ) from p(θ)p(y |θ), choosing ε essentially amounts to choosing the bandwidth of a KDE. There are some speciﬁc aspects that may deserve some investigation however: 1 The objective is to approximate a conditional density, that is p(θ|s(y )). (But approximating p(s(y )) may be interesting too.) 2 The marginal distribution of the simulated θ’s is known. 3 Could we use a bandwidth matrix instead? N. CHOPIN Intractability 10/ 54
11. 11. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Parametric interpretation of ABC It would be great to take s(y ) = y . In that way, the ABC posterior could be interpreted as the posterior distribution of the same model, but corrupted with noise (of size ε). See the following paper for a fast (EP) approximation of such an ABC posterior: Barthelm´, S. and Chopin, N. (2011). ABC-EP: Expectation e Propagation for Likelihood-free Bayesian Computation, ICML 2011, L. Getoor and T. Scheﬀer (eds), 289-296. (see also arXiv:1107.5959). N. CHOPIN Intractability 11/ 54
12. 12. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2ABC: summary We use ABC for very challenging models (generative/scientiﬁc models). We pay a heavy price for this: 1 First level of approximation is p(θ|y ) ≈ p(θ|s(y )) (althought not in ABC-EP). 2 Second level of approximation is p(θ|s(y )) ≈ πε (θ). 3 Huge CPU cost (but less in ABC-EP). 4 ABC-EP cannot be used in all situations. In the rest of the talk, we will deal with milder problems, and we will be able to avoid approximations. N. CHOPIN Intractability 12/ 54
13. 13. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Outline 1 Background 2 ABC methods for generative models 3 MC2 type methods 4 State-Space models, PMCMC 5 SMC2 N. CHOPIN Intractability 13/ 54
14. 14. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Basic framework Imagine a model such that p(y |θ) = p(x, y |θ) dx is intractable, but can be approximated by the following unbiased MC estimate: N 1 p(x j , y |θ) p (y |θ) = ˆ N qθ (x j ) j=1 where the x j ’s are N points sampled from the (user-chosen) proposal distribution qθ . N. CHOPIN Intractability 14/ 54
15. 15. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Naive question Can we simply replace p(y |θ) by p (y |θ)? i.e. ˆ MC2 From current point θn (plus p (y |θn ) from previous iteration) ˆ 1 Sample θp ∼ T (θn , dθp ) 2 Sample x 1:N ∼ qθp so as to compute p (y |θp ). ˆ 3 With probability 1 ∧ r , set θn+1 = θp , otherwise θn+1 = θn with p(θp )ˆ(y |θp )T (θp , θn ) p r= . p(θn )ˆ(y |θn )T (θn , θp ) p N. CHOPIN Intractability 15/ 54
16. 16. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Answer: yes, and the algorithm is exact! More precisely, this algorithm is a correct Metropolis step with respect to the following extended distribution:   N N j , y |θ) 1 p(x π(θ, x 1:N ) ∝ p(θ) qθ (x j )   N qθ (x j ) j=1 j=1 which is such that the marginal distribution of θ is precisely the true posterior distribution: π(θ, x 1:N ) dx 1:N = p(θ|y ). N. CHOPIN Intractability 16/ 54
17. 17. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Outline 1 Background 2 ABC methods for generative models 3 MC2 type methods 4 State-Space models, PMCMC 5 SMC2 N. CHOPIN Intractability 17/ 54
18. 18. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2State Space Models A system of equations Hidden states (Markov): p(x1 |θ) = µθ (x1 ) and for t ≥ 1 p(xt+1 |x1:t , θ) = p(xt+1 |xt , θ) = fθ (xt+1 |xt ) Observations: p(yt |y1:t−1 , x1:t−1 , θ) = p(yt |xt , θ) = gθ (yt |xt ) Parameter: θ ∈ Θ, prior p(θ). We observe y1:T = (y1 , . . . yT ), T might be large (≈ 104 ). x and θ will also be of several dimensions. There are several interesting models for which fθ cannot be written in closed form (but it can be simulated). N. CHOPIN Intractability 18/ 54
19. 19. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2State Space Models Some interesting distributions Bayesian inference focuses on: static: p(θ|y1:T ) dynamic: p(θ|y1:t ) , t ∈ 1 : T Filtering (traditionally) focuses on: ∀t ∈ [1, T ] pθ (xt |y1:t ) Smoothing (traditionally) focuses on: ∀t ∈ [1, T ] pθ (xt |y1:T ) N. CHOPIN Intractability 19/ 54
20. 20. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Examples Population growth model yt = xt + σw εt log xt+1 = log xt + b0 + b1 (xt )b2 + σ ηt θ = (b0 , b1 , b2 , σ , σW ). N. CHOPIN Intractability 20/ 54
21. 21. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Examples Stochastic Volatility (L´vy-driven models) e Observations (“log returns”): 1/2 yt = µ + βvt + vt t ,t ≥ 1 Hidden states (“actual volatility” - integrated process): k 1 vt+1 = (zt − zt+1 + ej ) λ j=1 N. CHOPIN Intractability 21/ 54
22. 22. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Examples . . . where the process zt is the “spot volatility”: k zt+1 = e −λ zt + e −λ(t+1−cj ) ej j=1 iid iid k ∼ Poi λξ 2 /ω 2 c1:k ∼ U(t, t + 1) ei:k ∼ Exp ξ/ω 2 The parameter is θ ∈ (µ, β, ξ, ω 2 , λ), and xt = (vt , zt ) . See the results N. CHOPIN Intractability 22/ 54
23. 23. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Why are those models challenging? . . . It is eﬀectively impossible to compute the likelihood p(y1:T |θ) = p(y1:T |x1:T , θ)p(x1:T |θ)dx1:T XT Similarly, all other inferential quantities are impossible to compute. N. CHOPIN Intractability 23/ 54
24. 24. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Problems with MCMC approaches Metropolis-Hastings: 1 p(θ|y1:T ) cannot be evaluated point-wise (marginal MH) 2 p(x1:T , θ|y1:T ) are high-dimensional and it is hard to design reasonable proposals Gibbs sampler (updates states and parameters): 1 The hidden states x1:T are typically very correlated and it is hard to update them eﬃciently in a block 2 Parameters and latent variables highly correlated Common: they are not designed to recover the whole sequence π(x1:t , θ | y1:t ) N. CHOPIN Intractability 24/ 54
25. 25. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Particle ﬁlters Consider the simpliﬁed problem of targeting pθ (xt+1 |y1:t+1 ) This sequence of distributions is approximated by a sequence of weighted particles which are properly weighted using importance sampling, mutated/propagated according to the system dynamics, and resampled to control the variance. Below we give a pseudo-code version. Any operation involving the superscript n must be understood as performed for n = 1 : Nx , where Nx is the total number of particles. N. CHOPIN Intractability 25/ 54
26. 26. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2 Step 1: At iteration t = 1, n (a) Sample x1 ∼ q1,θ (·). (b) Compute and normalise weights n n n n µθ (x1 )gθ (y1 |x1 ) n w1,θ (x1 ) w1,θ (x1 ) = n , W1,θ = N . q1,θ (x1 ) i i=1 w1,θ (x1 ) Step 2: At iteration t = 2 : T n 1:Nx (a) Sample the index at−1 ∼ M(Wt−1,θ ) of the ancestor an (b) Sample xtn ∼ qt,θ (·|xt−1 ). t−1 (c) Compute and normalise weights an an n at−1 fθ (xtn |xt−1 )gθ (yt |xtn ) t−1 wt,θ (xt−1 , xtn ) t−1wt,θ (xt−1 , xtn ) = an , n Wt,θ = i at−1 i qt,θ (xtn |xt−1 ) t−1 Nx i=1 wt,θ (xt−1 , xt ) N. CHOPIN Intractability 26/ 54
27. 27. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Particle ﬁltering time Figure: Three weighted trajectories x1:t at time t. N. CHOPIN Intractability 27/ 54
28. 28. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Particle ﬁltering time Figure: Three proposed trajectories x1:t+1 at time t + 1. N. CHOPIN Intractability 28/ 54
29. 29. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Particle ﬁltering time Figure: Three reweighted trajectories x1:t+1 at time t + 1 N. CHOPIN Intractability 29/ 54
30. 30. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Observations (i) (i) At each t, (wt , x1:t )Nx is a particle approximation of i=1 pθ (xt |y1:t ). Resampling to avoid degeneracy. If there were no interaction between particles there would be typically polynomial or worse increase in the variance of weights Taking qθ = fθ simpliﬁes weights, but mainly yields a feasible algorithm when fθ can only be simulated. N. CHOPIN Intractability 30/ 54
31. 31. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Unbiased likelihood estimator A by-product of PF output is that T Nx ˆ 1 (i) ZtN = wt Nx t=1 i=1 is an unbiased estimator of the likelihood Zt = p(y1:t |θ) for all t. Whereas consistency of the estimator is immediate to check, unbiasedness is subtle, see e.g Proposition 7.4.1 in Del Moral. The variance of this estimator grows typically linealy with T (and not exponentially) because of lack of independence. N. CHOPIN Intractability 31/ 54
32. 32. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2PSMC Breakthrough paper of Andrieu et al. (2011), based on the unbiasedness of the PF estimate of the likelihood. Marginal PMCMC From current point θn (and current PF estimate p (y |θn )): ˆ 1 Sample θp ∼ T (θn , dθp ) 2 Run a PF so as to obtain p (y |θp ), an unbiased estimate of ˆ p(y |θp ). 3 With probability 1 ∧ r , set θn+1 = θp , otherwise θn+1 = θn with p(θp )p(y |θp )T (θp , θn ) r= p(θn )ˆ(y |θn )T (θn , θp ) p N. CHOPIN Intractability 32/ 54
33. 33. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Outline 1 Background 2 ABC methods for generative models 3 MC2 type methods 4 State-Space models, PMCMC 5 SMC2 N. CHOPIN Intractability 33/ 54
34. 34. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Objectives 1 to derive sequentially p(θ, x1:t |y1:t ), p(y1:t ), for all t ∈ {1, . . . , T } 2 to obtain a black box algorithm (automatic calibration). N. CHOPIN Intractability 34/ 54
35. 35. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Main tools of our approach Particle ﬁlter algorithms for state-space models (this will be to estimate the likelihood, for a ﬁxed θ). Iterated Batch Importance Sampling for sequential Bayesian inference for parameters (this will be the theoretical algorithm we will try to approximate). Both are sequential Monte Carlo (SMC) methods N. CHOPIN Intractability 35/ 54
36. 36. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2IBIS SMC method for particle approximation of the sequence p(θ | y1:t ) for t = 1 : T . PF is not going to work here by just pretending that θ is a dynamic process with zero (or small) variance. Recall the path degeneracy problem. In the next slide we give the pseudo-code of the IBIS algorithm. Operations with superscript m must be understood as operations performed for all m ∈ 1 : Nθ , where Nθ is the total number of θ-particles. N. CHOPIN Intractability 36/ 54
37. 37. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Sample θm from p(θ) and set ω m ← 1. Then, at time t = 1, . . . , T (a) Compute the incremental weights and their weighted average Nθ 1 ut (θm ) = p(yt |y1:t−1 , θm ), Lt = Nθ × ω m ut (θm ), m m=1 ω m=1 (b) Update the importance weights, ω m ← ω m ut (θm ). (1) ˜ (c) If some degeneracy criterion is fulﬁlled, sample θm independently from the mixture distribution Nθ 1 Nθ ω m Kt (θm , ·) . m m=1 ω m=1 Finally, replace the current weighted particle system: ˜ (θm , ω m ) ← (θm , 1). N. CHOPIN Intractability 37/ 54
38. 38. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Observations Cost of lack of ergodicity in θ: the occasional MCMC move Still, in regular problems resampling happens at diminishing frequency (logarithmically) Kt is an MCMC kernel invariant wrt π(θ | y1:t ). Its parameters can be chosen using information from current population of θ-particles Lt is a MC estimator of the model evidence Infeasible to implement for state-space models: intractable incremental weights, and MCMC kernel N. CHOPIN Intractability 38/ 54
39. 39. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Our algorithm: SMC2 We provide a generic (black box) algorithm for recovering the sequence of parameter posterior distributions, but as well ﬁltering, smoothing and predictive. We give next a pseudo-code; the code seems to only track the parameter posteriors, but actually it does all other jobs. Superﬁcially, it looks an approximation of IBIS, but in fact it does not produce any systematic errors (unbiased MC). N. CHOPIN Intractability 39/ 54
40. 40. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Sample θm from p(θ) and set ω m ← 1. Then, at timet = 1, . . . , T , (a) For each particle θm , perform iteration t of the PF: If t = 1, sample independently x1 x ,m from ψ1,θm , and 1:N compute Nx 1 n,m p (y1 |θm ) = ˆ w1,θ (x1 ); Nx n=1 If t > 1, sample xt1:Nx ,m , at−1x ,m from ψt,θm 1:N 1:Nx ,m 1:Nx ,m conditional on x1:t−1 , a1:t−2 , and compute Nx 1 an,m ,m p (yt |y1:t−1 , θm ) = ˆ wt,θ (xt−1 , xtn,m ). t−1 Nx n=1 N. CHOPIN Intractability 39/ 54
41. 41. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2(b) Update the importance weights, ω m ← ω m p (yt |y1:t−1 , θm ) ˆ(c) If some degeneracy criterion is fulﬁlled, sample θm , x1:t x ,m , ˜1:t−1 independently from ˜ ˜1:N a1:Nx Nθ 1 Nθ ω m Kt θm , x1:t x ,m , a1:t−1 , · 1:N 1:Nx ,m m m=1 ω m=1 Finally, replace current weighted particle system: (θm , x1:t x ,m , a1:t−1 , ω m ) ← (θm , x1:t x ,m , ˜1:t−1 , 1) 1:N 1:Nx ,m ˜ ˜1:N a1:Nx ,m N. CHOPIN Intractability 40/ 54
42. 42. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Observations It appears as approximation to IBIS. For Nx = ∞ it is IBIS. However, no approximation is done whatsoever. This algorithm really samples from p(θ|y1:t ) and all other distributions of interest. One would expect an increase of MC variance over IBIS. The validity of algorithm is essentially based on two results: i) the particles are weighted due to unbiasedness of PF estimator of likelihood; ii) the MCMC kernel is appropriately constructed to maintain invariance wrt to an expanded distribution which admits those of interest as marginals; it is a Particle MCMC kernel. The algorithm does not suﬀer from the path degeneracy problem due to the MCMC updates N. CHOPIN Intractability 40/ 54
43. 43. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2The MCMC step ˜ ˜ ˜ (a) Sample θ from proposal kernel, θ ∼ T (θ, d θ). ˜ (b) Run a new PF for θ: sample independently 1:Nx 1:Nx (˜1:t , ˜1:t−1 ) from ψt,θ , and compute x a ˜ ˆt (θ, x 1:Nx , ˜1:Nx ). Z ˜˜ a 1:t 1:t−1 (c) Accept the move with probability ˜ ˆ ˜ ˜1:N a1:Nx ˜ p(θ)Zt (θ, x1:t x , ˜1:t−1 )T (θ, θ) 1∧ . ˆ ˜ p(θ)Zt (θ, x 1:Nx , a1:Nx )T (θ, θ) 1:t 1:t−1 It can be shown that this is a standard Hastings-Metropolis kernel with proposal ˜ ˜1:N a1:N ˜ 1:N a1:N qθ (θ, x1:t x , ˜1:t x ) = T (θ, θ)ψt,θ (˜1:t x , ˜1:t x ) ˜ x 1:N 1:Nx invariant wrt to an extended distribution πt (θ, x1:t x , a1:t−1 ). N. CHOPIN Intractability 41/ 54
44. 44. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Some advantages of the algorithm Immediate estimates of ﬁltering and predictive distributions Immediate and sequential estimator of model evidence Easy recovery of smoothing distributions Principled framework for automatic calibration of Nx Population Monte Carlo advantages N. CHOPIN Intractability 42/ 54
45. 45. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Numerical illustrations: SV 1.0 800 700 8 0.8 600 Squared observations Acceptance rates 6 0.6 500 Nx 400 4 0.4 300 2 0.2 200 100 0 0.0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000 Time Iterations Iterations (a) (b) (c) Figure: Squared observations (synthetic data set), acceptance rates, and illustration of the automatic increase of Nx . See the model N. CHOPIN Intractability 43/ 54
46. 46. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Numerical illustrations: SV T = 250 T = 500 T = 750 T = 1000 8 6 Density 4 2 0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 1.0 µ Figure: Concentration of the posterior distribution for parameter µ. N. CHOPIN Intractability 44/ 54
47. 47. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Numerical illustrations: SV Multifactor model k1 k2 1/2 yt = µ+βvt +vt t +ρ1 e1,j +ρ2 e2,j −ξ(w ρ1 λ1 +(1−w )ρ2 λ2 ) j=1 j=1 where vt = v1,t + v2,t , and (vi , zi )i=1,2 are following the same dynamics with parameters (wi ξ, wi ω 2 , λi ) and w1 = w , w2 = 1 − w . N. CHOPIN Intractability 45/ 54
48. 48. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Numerical illustrations: SV Evidence compared to the one factor model variable 20 Multi factor without leverage 4 Multi factor with leverage Squared observations 15 2 10 0 5 −2 100 200 300 400 500 600 700 100 200 300 400 500 600 700 Time Iterations (a) (b) Figure: S&P500 squared observations, and log-evidence comparison between models (relative to the one-factor model). N. CHOPIN Intractability 46/ 54
49. 49. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Final Remarks A powerful framework A generic algorithm for sequential estimation and state inference in state space models: only requirements are to be able (a) to simulate the Markov transition fθ (xt |xt−1 ), and (b) to evaluate the likelihood term gθ (yt |xt ). The article is available on arXiv and our web pages A package is available at: http://code.google.com/p/py-smc2/. N. CHOPIN Intractability 47/ 54
50. 50. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Appendix N. CHOPIN Intractability 48/ 54
51. 51. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Why does it work? - Intuition for t = 1 At time t = 1, the algorithm generates variables θm from the prior p(θ), and for each θm , the algorithm generates vectors x1 x ,m of 1:N 1:N particles, from ψ1,θm (x1 x ). N. CHOPIN Intractability 49/ 54
52. 52. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Thus, the sampling space is Θ × X Nx , and the actual “particles” ofthe algorithm are Nθ independent and identically distributed copies 1:Nof the random variable (θ, x1 x ), with density: Nx 1:N n p(θ)ψ1,θ (x1 x ) = p(θ) q1,θ (x1 ). n=1 N. CHOPIN Intractability 50/ 54
53. 53. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Then, these particles are assigned importance weightscorresponding to the incremental weight functionˆ 1:N −1 Nx nZ1 (θ, x1 x ) = Nx n=1 w1,θ (x1 ).This means that, at iteration 1, the target distribution of thealgorithm should be deﬁned as: 1:N 1:N ˆ 1:N Z1 (θ, x1 x ) π1 (θ, x1 x ) = p(θ)ψ1,θ (x1 x ) × , p(y1 )where the normalising constant p(y1 ) is easily deduced from the ˆ 1:Nproperty that Z1 (θ, x1 x ) is an unbiased estimator of p(y1 |θ). N. CHOPIN Intractability 51/ 54
54. 54. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2Direct substitutions yield Nx n Nx n 1:N p(θ) i µθ (x1 )gθ (y1 |x1 ) 1π1 (θ, x1 x ) = q1,θ (x1 ) n p(y1 ) Nx q1,θ (x1 ) i=1 n=1   N  Nx 1 x p(θ) n n i  = µθ (x1 )gθ (y1 |x1 ) q1,θ (x1 ) Nx p(y1 )   n=1 i=1,i=nand noting that, for the triplet (θ, x1 , y1 ) of random variables, p(θ)µθ (x1 )gθ (y1 |x1 ) = p(θ, x1 , y1 ) = p(y1 )p(θ|y1 )p(x1 |y1 , θ)one ﬁnally gets that:   Nx Nx 1:N p(θ|y1 ) n  i  π1 (θ, x1 x ) = p(x1 |y1 , θ) q1,θ (x1 ) . Nx   n=1 i=1,i=n N. CHOPIN Intractability 52/ 54
55. 55. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2By a simple induction, one sees that the target density πt atiteration t ≥ 2 should be deﬁned as: ˆ 1:N 1:Nx Zt (θ, x1:t x , a1:t−1 ) 1:N 1:Nx 1:N 1:Nxπt (θ, x1:t x , a1:t−1 ) = p(θ)ψt,θ (x1:t x , a1:t−1 ) × p(y1:t )and the following Proposition N. CHOPIN Intractability 53/ 54
56. 56. Background ABC methods for generative models MC2 type methods State-Space models, PMCMC SMC2PropositionThe probability density πt may be written as: 1:N 1:Nx πt (θ, x1:t x , a1:t−1 ) = p(θ|y1:t )   N  N  1 x p(xn |θ, y1:t )  x    1:t i × t−1 q1,θ (x1 ) Nx Nx   n=1  i=1  n   i=ht (1)    t  Nx    i as−1 i  i as−1 × Ws−1,θ qs,θ (xs |xs−1 )  s=2 i=1    n  i=ht (s) N. CHOPIN Intractability 54/ 54