Approximate Inference: Markov Chain Monte Carlo

1
Approximate Inference Markov Chain Monte Carlo

• by
• Gary Holness

2
Outline

• Summarize 50 years of work!!
• Touch on ideas of the theory
• Motivate language with classical stats
• Bayesian improvement
• Sampling techniques
• Mention Markov Chain
• Metropolis-Hastings framework
• -general
• -simulated annealing, Gibbs sampling

3
Distributions govern data

• Distribution is a lawf(x q )
• x (x1,,xn)q (q1,, qn)q (m , s)
• Indexed by parameters
• Knowing params, can make statements about yet
  unseen events

4
Point estimation
L(q)
q
search parameter space

• Estimator q is statistic T t(X1,,Xn) can
  substitute for q
• Consider pdf as some f(q) over X L(q
  X1,,Xn ) with our random sample
• q argmaxq L(q X1,,Xn ) MLE

5
Bayesian improvement

• Point estimation gives single value
• Treat unknown quantity as random variable
• Have generator f(xq) of evidence
• Since q is r.v. write pdf as a conditional f(xq)
• Choose prior dist for r.v. p(q)
• Using Bayes rule compute posterior

6
Bayes contd

• Given posterior for q we can talk about -
  expectations of arbitrary functions g(q)
  - mode (MAP est.) qMAP argmaxq p(qx)
• Unsure about q - p(q) high variance -
  improper prior
• Posterior not well known - conjugate choose
  prior so posterior is same family

7
What else is hard

• The proportionality const - only know p(qx)
  within constant - large discrete spaces,
  marginalizing bad O(an)
• In Bayes net
• Suppose have posterior for x p(xe)
• Computing E xg(x) means integrating/summing
  over large space
• OK, use sampling can be hard to generate
  samples from posterior

8
Monte Carlo (MC)

• Need cheaper numerical integration
• MC techniques originally for particle physics
• Use sampling to approximate - let Y f(X)
  - let X1,,Xn be r.s. Xip(xe) r.s. ?
  I.I.D -
• Stillgenerating r.s. from support of p(xe)
  hard
• - sample where p(xe) has mass - sample
  from scratch every time

9
What if its hard to generate samples?

• Propose another dist easier to generate samples
• Compute estimator for g(x) using samples
  influence the target dist b.
• Amount the target and proposal agree on mass of
  support importance weight

10
Proposal Dist Example Importance Sampling

• We want to estimate function g(x)
• Let f(x) be posterior dist
• Let q(x) be proposal dist
• Sample X1Xn where Xiq(x)

11
Sampling themes

• Numerical Integration
• Representative sample from support
• Approximation for inference
• Difficult distributions
• Proposal distribution
• Account for discrepancies where p and q put mass
• Talk in terms of improvements

12
MC direct sampling

• Sample from priors and cond in topological order
• Sample from posterior
• Compute prob of event by counting samples (in the
  limit)
• Hard to sample from posterior

13
MC rejection sampling

• Better estimation
• Sample from priors/conditionals, P, given in
  Bayes net
• Reject samples which disagree with evidence by
  direct filtering
• Or sample from simpler proposal distribution q

14
MC rejection sampling

• P(x) is
• Compare mass assigned to sample by P,Q massx ?
  Q gt massx?P
• Compute posterior from accepted samples
• Problem choosing the right Q Q very
  different from P ? reject many samples

15
MC Weighted Sampling

• Generate samples consistent w/evidence
• P(Rain Sprinkler true, WetGrass true)
• Fix evidence vars

16
Weighted sampling contd

• sample non-evidence vars (using cond tables)
• Count sampled events matching query
• Samples credibility modulated by
• Only incorporates evidence vars which influence
  the particular sample

17
MC Weighted sampling

• Each sampled event gets weight
• Compute posterior by weighted sum of sampled
  events
• Credibility ? influence on probability mass

18
MCMC

• WS-MC performance degrades with increasing number
  of evidence variables
• Large spaces individual samples carry small
  weight
• Weighted still generates samples from scratch
• Why not generate sample by small modification to
  known event?

19
Markov Chain

• Given an n-dimensional state space
• Random vector x (x1,,xn)
• x(t) x at time-step t
• x(t) transitions to x(t1) with probP(x(t)
  x(t-1),,x(1)) T(x(t) x(t-1))
• Homogenous chain determined by state x, fixed
  transition kernel T (rows sum to 1)

20
Markov chain

• Irreducible transition graph connected
• Aperiodic not trapped in cycles
• Detailed bal prob(x(i)?x(i-1))
  prob(x(I-1)?x(i))
• Detailed bal ? stationary dist exists

21
Metropolis-Hastings

• Treat target dist as stationary distribution
• Build transition matrix so that we satisfy
  detailed balance
• While sampling from easier proposal dist

22
MCMC Metropolis-Hastings

• Have invariant dist p(x)
• Proposal dist q(x x)
• Sample x given x from q(xx)
• Chain transitions to state x with probability
• prob(x?x) prob(x?x)

23
Metropolis Hastings

• Our transition Kernel becomes

24
Metropolis-Hastings

• Initialize x(0)
• For i 0 to N-1 - sample uUnif(0,1) -
  sample x q(xx) - if u lt
• x(i1) x //transition
  else x(i1) x(i) //stay in curr state

25
General MCMC example (Jordan)

• q(xx) N(x(i),100)
• p(x) 0.3 exp(-0.2x2) 0.7 exp(-0.2(x-10)2)

26
Metropolis-Hastings

• Random walk through state space (jumps)
• Can simulate multiple chains in parallel
• Much hinges on proposal dist, q
• Want to visit state space where p(x) puts mass
• Modes of p(x) visited while A(x,x) high
• Chain mixes well

27
Simulated Annaeling Global Optimization

• Given p(x) want to find global max
• Compute MAP estimate if p posterior
• simulate by markov chain to get p(x)

28
Simulated Annealing

• Search inefficient, few samples from p(x)s mode
• Simulate non-homogeneous chain (transition kernel
  changes)
• Invariant dist p changes with temperature
• concentrates on global maxima

29
MCMC Simulated Annealing

• Initialize x(0)
• For i 0 to N-1 - sample uUnif(0,1) -
  sample x q(xx) - if u lt
• x(i1) x else
  x(i1) x(i)
• -set Ti1 on temperature schedule

30
MCMC-SA

• Must choose appropriate schedule
• Appropriate proposal dist
• Shown to converge for Ti C (ln(i T0)-1)Geman
  1984, Van Laarhoven 1987
• C and T0 problem dependent
• What if state space really huge?

31
Blocking

• Large state space state vector comprised of many
  components (high dimension)
• Some components can be correlated
• Sample components one at a time
• Each block has own transition kernel governing
  how it is changed
• Blocks can be correlated components
• Can use blocks to get better mixing

32
Markov chain w/cycle of Kernels

• Sample from joint dist XX1,,Xn
• With prob P(x1,,xn)
• Build chain from kernel Tk for k1nTk(x(i1)
  x(i)) P(x(i1)k x-k)
• x-k xi i !k
• Using kernel Tk change component k by sampling
• Apply Tk for k1n in sequence (or rand)

33
Cycle of kernels

• T(x(i1) x(i)) Tj(x(j1) x(j))
• T T1T2Tn
• Sample from proposal distribution for each
  component q(xx(i))
• Acceptance probability still
• Play with blocking/cycling to improve mixing

34
Gibbs Sampling

• Initialize x
• For i 0 to N-1 - sample x1(i1) - sample
  x2(I1) - sample xn(i1) - sample
  uUnif(0,1) - if u lt A(x(i),x) by choosing
  q p, A1 x(i1) x else
  x(i1) x(i)

35
Gibbs sampling in Bayes net

• Dont sample from full component-at-a-time
  conditional
• Use Structure of Bayes net
• Condition on Markov Blanket
• A,B,C,I,L ? G Mb(G)

36
Summary

• Large problem approximate by sampling
• Difficult distribution Use proposal
• Markov Chain sampling by re-use/walk
• Detailed balance target dist exists
• Sample and test against acceptance
• Want good mixing
• Many, many, many, MCMC variations out there

37
Thank you for listeningHow can MCMC make all
your dreams come true?