Jay : Seaman Iris Last Lecture : Importance Sampling Xsnqcx - PowerPoint PPT Presentation

Lecture Markov Monte to Carlo Chain : De Young Scribes Jay : Seaman Iris

Last Lecture : Importance Sampling Xsnqcx Generate from Idea samples ) : distribution similar that proposal to ) Mk a is weight MCN high : y underrepresented gcxi E ""iµq%*+ use :-. . overrepresented ✓ proposals by x fdxtrcxlfki Elftlx ) ) = = = xsngcxl = =

Likelihood Weighting Default Choice fan : Bayes Net Assume Proposal = ply ) 2- ycx ) = = = Prior Proposal Set to l x ) q = Likelihood Weights Importance : = & Is f Ks ) us = F- s , = [ W S - i - ' s = )

Problem Models Motivating Hidden Markov : §→€→ Yt yi t & Ze 7 z . , , ← It Posterior Goal Parameters : an O I y ) PC = t libel ? Will ihocd weighty wah I ' O w =

Markov Chain Monte Carlo S I Use - Idea to previous sample propose : x the ' sample neat x A of Masher Chain variables random : sequence " ,Xs X when Markov chain ) I discrete-time is a . . . , " xslxs ' ' - Ix ,x , . . . ) 1×51×5-1 - ' ' is ( x' ) p I x = p Markov A Chain homogenous when is ' ) pcxs.es/Xs-!xs-i)=pCX'=xslX-xs .

Manha Monte Carlo Chain A Manha Convergence chain to converges : density target when ( 17 ) x a p(Xs=x lying ) n(X=× ) = . "II÷u III : 2 *¥*y¥¥± . which X=x in visited is , " frequency " with h(X=x ) J

Markov Monte Carlo Chain Balance A Markov Detailed chain homogenous : balance detailed satisfies when ' ) plxcx 't I X' Ix ) MCX ) Mix p = invariant leaves Implication ' 1×1 pcx Mix ) i x ) 17 C = = = ' start If with sample next x a you n ' ) then then ' sample pl Mex ) and XIX XIX ~ x n

Metropolis Hastings - from Starting xs sample Idea current the : = nain accept proposal qcxixs ) ' generate and n a x ' xst probability with ' × = . ) ( I a , probability proposal with reject the C ) o I - "=xs the xs retain sample and previous

Metropolis Hastings - from Starting xs sample Idea current the : = mmin accept proposal qcxixs ) ' generate and ~ a × '=x xst probability with ' ' ) ' ) M q( ( XIX . X ) ) ( l a i qkllx ( × ) n probability with ( proposal reject the ) l d - "=xs xs the retain sample previous and Show Markov Exercise the that chain : balance xs ' detailed satisfies x . . .

- Hastings Balance Detailed etropolis : Balance Detailed : ' ) M ( I ' Ix ) X ) pcxcx 't Mix X = p - Hastings Metropolis Define : kcx Kix = minfl ' IN a = , 7457,44272¥ ' Ix ) MCA = min ( ) + , =

- Hastings Densities Unrormalired etropolis : Nice Can calculate property prob acceptance : ' ) from ✓ C x ) densities unhormaliced and jcx i mix mmin ' ) ' I 17 ( 9 ( XIX . x ) ) ( I a = qcx ' Ix , nain - ) ( I = , main 81×1 ) = ( I = ,

- Hastings Metropolis Choosing proposals i Sample from Mtt Independent proposers : pretor - ( IX ) ' ) ' sample ) C X Independent from 9 pcx = prau . in ) ( I a mm = , in mind ) ) ( I mm - = = , , but low Straightforward acceptance prob , ( )

and ( " dark of - Hastings Metropolis Choosing proposals : 2 Gaussian variables Continuous : Norm ( 82 ) • ' ' ) qcx Ix x x ; . = , ' small or far proposal Trade off van , - proposal Symmetric 82 small prob too good acceptance A - ; , 91×4×1=941×1 ) but high correlation between samples min ( ) ← I less I 82 but correlation large too - : , min ( lower prob ) acceptance A I = - , Rule of ' thumb of time I 0.234 to make 9 ,

Lecture ) Gibbs ( Next Sampling Propose Idea 1 time at variable : a , holding other variables constant ) C ply y x pix ,Xa ) ) Ix Xz = , , . X , / ' X. ply Xa ) ply xi n = , , , xi - Acceptance Ratio Car accept prob I with : = min ( ) A I = , , I =

MCMC Importance Sampling vs . , O ) IZ O ) 2- ply ) ) 8107 ply NCO y ( = = = Sampling importance i - P' {j% Anglo Elws ) ws on geo ) pigs = - - , a y I check Gives of Guess and estimate marginal - Hastings Metropolis = { " , - Unit ( 0,1 ) Os " '1o u > a u ps min ( 8105910541014 " ) 9104054 , ) O ca , n g = , " HO " but " estimate of marginal climbing hill Can do no

Marginal Likelihoods Computing Motivation Model comparison : ' K ? Question ' How clusters : many * Low 109 High if ) ply ply

Marginal Likelihoods Computing Motivation Model : ' K ? Question ' How clusters comparison : many * Low 109 High if ) ply ply Fewer bad bad Lots of 0 O

Marginal Likelihoods Computing Motivation Model : ' K ? Question ' How clusters : many * Low 109 High if ) ply ply comparison Fewer bad bad Lots of 0 0 Bayesian Approach : likelihoods marginal Compare angmax / * K plylk ) do angmax = = " } K km ke { I . . . , ,

Annealed Importance Sampling intermediate Idea Sample from target by of yco ) r.mg I : way distributions , I O ) lol yn ya Idea Use MLMC to generate proposals 2 :

Annealed Importance Sampling O @ . . :-. 4%0%4 . ) - got Oi Initialization w Onsnkn ,( I On ! ) On wins Transition = - . )