Monte-Carlo
  • Introduction
  • AIS
  • Generate R.V.
    • Special Distribution
    • Copulas
    • Minimum of Two Exponential
  • Gibbs
    • Comparing two groups
    • Linear Regression
    • Simulation of Exp-Abs-xy
  • Markov Chain
  • MC Approximation
  • MC Integration
    • Rao-Blackwellization
  • MC Optimization
  • MCMC
    • MCMC diagnostics
  • Metropolis-Hastings
    • Metropolis
    • Independent MH
    • Random Walk MH
    • ARMS MH
  • PBMCMC
  • RJMCMC
  • Diagnosing Convergence
  • SMCTC
  • Tempering
    • Parallel Tempering
  • Misc
    • R vs. Julia
  • References
Powered by GitBook
On this page
  • One Simple Way
  • One Example

AIS

PreviousIntroductionNextGenerate R.V.

Last updated 6 years ago

One Simple Way

  1. start with a trial density, say g0(x)=tα(x;μ0,Σ0)g_0(x) = t_\alpha (x; \mu_0,\Sigma_0)g0​(x)=tα​(x;μ0​,Σ0​)

  2. with weighted Monte Carlo samples, estimate the parameters, mean and covariate matrix, and construct new trial density, say g1(x)=tα(x;μ1,Σ1)g_1(x) = t_\alpha (x; \mu_1,\Sigma_1)g1​(x)=tα​(x;μ1​,Σ1​)

  3. construct a certain measure of discrepancy between the trial distribution and the target distribution, such as the coefficient of variation of importance weights, does not improve any more.

One Example

Implement an Adaptive Importance Sampling algorithm to evaluate mean and variance of a density

π(x)∝N(x;0,2I4)+2N(x;3e,I4)+1.5N(x;−3e,D4)\pi(\mathbf{x})\propto N(\mathbf{x;0}, 2I_4) + 2N(\mathbf{x; 3e}, I_4) + 1.5 N(\mathbf{x; -3e}, D_4)π(x)∝N(x;0,2I4​)+2N(x;3e,I4​)+1.5N(x;−3e,D4​)

where e=(1,1,1,1),I4=diag(1,1,1,1),D4=diag(2,1,1,.5)\mathbf{e} = (1,1,1,1), I_4 = diag(1,1,1,1), D_4 = diag(2,1,1,.5)e=(1,1,1,1),I4​=diag(1,1,1,1),D4​=diag(2,1,1,.5).

A possible procedure is as follows:

  1. start with a trial density g0=tν(0,Σ)g_0 =t_{\nu}(0, \Sigma)g0​=tν​(0,Σ)

  2. Recursively, build

    gk(x)=(1−ϵ)gk−1(x)+ϵtν(μ,Σ)g_k(\mathbf{x})=(1-\epsilon)g_{k-1}(\mathbf {x}) + \epsilon t_{\nu}(\mu, \Sigma)gk​(x)=(1−ϵ)gk−1​(x)+ϵtν​(μ,Σ)

    in which one chooses (ϵ,μ,Σ)(\epsilon, \mu, \Sigma)(ϵ,μ,Σ) to minimize the variation of coefficient of the importance weights.