Introduction

This package provides implementations of some general-purpose random-walk based adaptive MCMC algorithms, including the following:

The aim of the package is to provide a simple and modular general-purpose implementation, which may be easily used to sample from a log-target density, but also used in a variety of custom settings.

See also AdaptiveParticleMCMC.jl which uses this package with SequentialMonteCarlo.jl for adaptive particle MCMC.

Installation

To get the latest registered version:

using Pkg
Pkg.add("AdaptiveMCMC")

To install the latest development version:

using Pkg
Pkg.add(url="https://github.com/mvihola/AdaptiveMCMC.jl")

Sampling from log-posteriors

The package provides an easy-to-use adaptive random-walk Metropolis sampler, which samples (in principle) from any probability distribution $p$, whose log-density values can be evaluated point-wise.

# Load the package
using AdaptiveMCMC

# Define a function which returns log-density values:
log_p(x) = -.5*sum(x.^2)

# Run 10k iterations of the Adaptive Metropolis:
out = adaptive_rwm(zeros(2), log_p, 10_000; algorithm=:am)

# Calculate '95% credible intervals':
using Statistics
mapslices(x->"$(mean(x)) ± $(1.96std(x))", out.X, dims=2)

See Adaptation state for explanation of the different algorithm options:

  • :am = AdaptiveMetropolis
  • :ram = RobustAdaptiveMetropolis
  • :asm = AdaptiveScalingMetropolis
  • :aswam = AdaptiveScalingWithinAdaptiveMetropolis

There are a number of other optional keyword arguments, too:

AdaptiveMCMC.adaptive_rwmFunction
out = adaptive_rwm(x0, log_p, n; kwargs)

Generic adaptive random walk Metropolis algorithm from initial state vector x0 targetting log probability density log_p run for n iterations, including adaptive parallel tempering.

Arguments

  • x0::Vector{<:AbstractFloat}: The initial state vector
  • log_p::Function: Function that returns log probability density values (up to an additive constant) for any state vector.
  • n::Int: Total number of iterations

Keyword arguments

  • algorithm::Symbol: The random walk adaptation algorithm; current choices are :ram (default), :am, :asm, :aswam and :rwm. (Alternatively, if algorithm is a vector of AdaptState, then this will be used as an initial state for adaptation.)
  • b::Int: Burn-in length: b:th sample is the first saved sample. Default ⌊n/5⌋
  • thin::Int: Thinning factor; only every thin:th sample is stored; default 1
  • fulladapt::Bool: Whether to adapt after burn-in; default true
  • Sp: Saved adaptive state from output to restart MCMC; default nothing
  • Rp: Saved rng state from output to restart MCMC; default nothing
  • indp::Int: Index of saved adaptive state to restart MCMC; default 0
  • rng::AbstractRNG: Random number generator; default Random.GLOBAL_RNG
  • q::Function: Zero-mean symmetric proposal generator (with arguments x and rng); default q=randn!(x, rng)
  • L::Int: Number of parallel tempering levels
  • acc_sw::AbstractFloat: Desired acceptance rate between level swaps; default 0.234
  • all_levels::Bool: Whether to store output of all levels; default false
  • log_pr::Function: Log-prior density function; default log_pr(x) = 0.0.
  • swaps::Symbol: Swap strategy, one of: :single (default, single randomly picked swap) :randperm (swap in random order) :sweep (up- or downward sweep, picked at random) :nonrev (alternate even/odd sites as in Syed, Bouchard-Côté, Deligiannidis, Doucet, arXiv:1905.02939)
  • progress::Union{Bool,Progress}: Whether a progress meter is shown; default false

Note that if log_pr is supplied, then log_p(x) is regarded as the log-likelihood (or, equivalently, log-target is log_p(x) + log_pr(x)). Tempering is only applied to log_p, not to log_pr.

The output out.X contains the simulated samples (column vectors).out.allX[k]fork>=2` contain higher temperature auxiliary chains (if requested)

Examples

log_p(x) = -.5*sum(x.^2)
o = adaptive_rwm(zeros(2), log_p, 10_000; algorithm=:am)
using MCMCChains, StatsPlots # Assuming MCMCChains & StatsPlots are installed...
c = Chains(o.X[1]', start=o.params.b, thin=o.params.thin); plot(c)
source

With adaptive parallel tempering

If the keyword argument L is greater than one, then the adaptive parallel tempering algorithm (APT) of Miasojedow, Moulines & Vihola (2013) is used. This can greatly improve mixing with multimodal distributions.

Here is a simple multimodal distribution sampled with normal adaptive random walk Metropolis, and with APT:

# Multimodal target of dimension d.
function multimodalTarget(d::Int, sigma2=0.1^2, sigman=sigma2)
    # The means of mixtures
    m = [2.18 5.76; 3.25 3.47; 5.41 2.65; 4.93 1.50; 8.67 9.59;
         1.70 0.50; 2.70 7.88; 1.83 0.09; 4.24 8.48; 4.59 5.60;
         4.98 3.70; 2.26 0.31; 8.41 1.68; 6.91 5.81; 1.14 2.39;
         5.54 6.86; 3.93 8.82; 6.87 5.40; 8.33 9.50; 1.69 8.11]'
    n_m = size(m,2)
    @assert d>=2 "Dimension should be >= 2"
    let m=m, n_m=size(m,2), d=d
        function log_p(x::Vector{Float64})
            l_dens = -0.5*(mapslices(sum, (m.-x[1:2]).^2, dims=1)/sigma2)
            if d>2
                l_dens .-= 0.5*mapslices(sum, x[3:d].^2, dims=1)/sigman
            end
            l_max = maximum(l_dens) # Prevent underflow by log-sum trick
            l_max + log(sum(exp.(l_dens.-l_max)))
        end
    end
end

using AdaptiveMCMC
n = 100_000; L = 2
rwm = adaptive_rwm(zeros(2), multimodalTarget(2), n; thin=10)
apt = adaptive_rwm(zeros(2), multimodalTarget(2), div(n,L); L = L, thin=10)

# Assuming you have 'Plots' installed:
using Plots
plot(scatter(rwm.X[1,:], rwm.X[2,:], title="w/o tempering", legend=:none),
scatter(apt.X[1,:], apt.X[2,:], title="w/ tempering", legend=:none), layout=(1,2))

What the APT is actually based on? Parallel tempering is a MCMC algorithm which samples from a product density proporitional to:

\[\prod_{i=1}^L p^{\beta(i)}(x^{(i)}),\]

where (the 'inverse temperatures') $1 = \beta(1) > \beta(2) > \cdots > \beta(L) > 0$.

In the end, the 'first level' is of interest (and samples of the first level are usually used for estimation), whereas the tempered levels $i=2,\ldots,L$ are auxiliary, which help the sampler to move between modes of a multi-modal target. The easier moving is because the tempered densities $p^{\beta(i)}$ are 'flatter' than $p$ for any $\beta(i)<1$.

The sampler consists of two types of MCMC moves:

  • Independent adaptive random-walk Metropolis moves on individual levels $i$, targetting tempered densities $p^{\beta(i)}$.
  • Switch moves, where swaps of adjacent levels $x^{(i)} \leftrightarrow x^{(i+1)}$ are proposed, and the moves are accepted with (Metropolis-Hastings) probability $\min\big\{1, \frac{p^{\beta(i)-\beta(i+1)}(x^{(i+1)})}{p^{\beta(i)-\beta(i+1)}(x^{(i)})}\big\}$.

In the APT, each random walk sampler for each individual level is adapted totally independently, following exactly the same mechanism as before. Additionally, the APT adapts the inverse temperatures $\beta(2),\ldots,\beta(L)$, in order to reach the average switch probability $0.234$. More precisely, adaptation mechanism tunes the parameters $\rho^{(i)}$, which determine

\[\frac{1}{\beta^{(i)}} = \frac{1}{\beta^{(i-1)}} + e^{\rho^{(i)}},\]

and the adaptation is similar to Adaptive scaling Metropolis: if swap $x^{(i-1)}\leftrightarrow x^{(i)}$ is proposed the $k$:th time, the parameter is updated as follows:

\[\rho_k^{(i)} = \rho_{k-1}^{(i)} + \gamma_k (\alpha_k^{(\text{swap }i)} - \alpha_*),\]

where $\alpha_k^{(\text{swap }i)}$ is the swap probability.

Restarting simulation

Simulation can be restarted, or continued after one simulation. Here is an example:

using AdaptiveMCMC, Random
log_p(x) = -.5*sum(x.^2)
Random.seed!(12345)
# Simulate 200 iterations first:
out = adaptive_rwm(zeros(2), log_p, 200)
# Simulate 100 iterations more:
out2 = adaptive_rwm(out.X[:,end], log_p, 100; Sp=out.S, Rp=out.R, indp=200)
# This results in exactly the same output as simulating 300 samples in one go:
Random.seed!(12345)
out2_ = adaptive_rwm(zeros(2), log_p, 300)

Using individual modules in a custom setting

In many cases, the simple samplers provided by adaptive_rwm are not sufficient, but a custmised sampler is necessary. For instance:

  • Adaptative sampler is used only for certain parameters, whilst others are updated by another MCMC scheme, such as with Gibbs moves.
  • The sampler state is large, and simulations cannot be saved in memory.

The package provides simple building blocks which allow such custom scenarios. Here is a simple example how the individual components can be used:

using AdaptiveMCMC

# Sampler in R^d
function mySampler(log_p, n, x0)

    # Initialise random walk sampler state: r.x current state, r.y proposal
    r = RWMState(x0)

    # Initialise Adaptive Metropolis state (with default parameters)
    s = AdaptiveMetropolis(x0)
    # Other adaptations are: AdaptiveScalingMetropolis,
    # AdaptiveScalingWithinAdaptiveMetropolis, and RobustAdaptiveMetropolis

    X = zeros(eltype(x0), length(x0), n) # Allocate output storage
    p_x = log_p(r.x)                     # = log_p(x0); the initial log target
    for k = 1:n

        # Draw new proposal r.x -> r.y:
        draw!(r, s)

        p_y = log_p(r.y)                      # Calculate log target at proposal
        alpha = min(one(p_x), exp(p_y - p_x)) # The Metropolis acceptance probability

        if rand() <= alpha
            p_x = p_y

            # This 'accepts', or interchanges r.x <-> r.y:
            # (NB: do not do r.x = r.y; these are (pointers to) vectors!)
            accept!(r)
        end

        # Do the adaptation update:
        adapt!(s, r, alpha, k)

        X[:,k] = r.x   # Save the current sample
     end
    X
end

# Standard normal target for testing
normal_log_p(x) = -mapreduce(e->e*e, +, x)/2

# Run 1M iterations of the sampler targetting 30d standard Normal:
X = mySampler(normal_log_p, 1_000_000, zeros(30))

See Random-walk sampler state and Adaptation state for more details about these components.