more flexible proposal options

src/crossover.jl

@@ -7,7 +7,7 @@ Performs crossover step for each particle pt in the chain
 function crossover!(model, de, group)
     for pt in group
         # generate the proposal
-        proposal = generate_proposal(de, pt, group)
+        proposal = de.generate_proposal(de, pt, group)
         # compute the weight of the proposal: prior loglikelihood + data loglikelihood
         compute_posterior!(de, model, proposal)
         # accept proposal according to Metropolis-Hastings rule
@@ -22,8 +22,7 @@ Generate proposal according to θ' = θt + γ1(θm − θn) + γ2(θb − θt) +
 * `Pt`: current particle
 * `group`: a group of particles
 """
-# 2.38/sqrt(2d)
-function generate_proposal(de, Pt, group)
+function random_gamma(de, Pt, group)
     Np = length(Pt.Θ)
     # select the base particle θb
     Pb = select_base(group)
@@ -34,7 +33,7 @@ function generate_proposal(de, Pt, group)
     # sample gamma weights
     γ₁ = rand(Uniform(.5, 1))
     # set γ₂ = 0 after burnin
-    de.iter > de.burnin ? γ₂ = 0.0 : γ₂ = rand(Uniform(.5, 1))
+    γ₂ = de.iter > de.burnin ? 0.0 : rand(Uniform(.5, 1))
     # sample noise for each parameter
     b = Uniform(-de.ϵ, de.ϵ)
     # compute proposal value
@@ -44,15 +43,66 @@ function generate_proposal(de, Pt, group)
     return Θp
 end
 
+"""
+Generate proposal according to θ' = θt + γ(θm − θn) + b
+where γ = 2.38
+* `de`: differential evolution object
+* `Pt`: current particle
+* `group`: a group of particles
+"""
+function fixed_gamma(de, Pt, group)
+    Np = length(Pt.Θ)
+    # select the base particle θb
+    Pb = select_base(group)
+    # group without Pb
+    group_diff = setdiff(group, [Pt])
+    # sample particles for θm and θn
+    Pm,Pn = sample(group_diff, 2, replace=false)
+    # sample gamma weights
+    γ = 2.38
+    # sample noise for each parameter
+    b = Uniform(-de.ϵ, de.ϵ)
+    # compute proposal value
+    Θp = Pt + γ * (Pm - Pn) + b
+    # reset each parameter to previous value with probability (1-κ)
+    recombination!(de, Pt, Θp)
+    return Θp
+end
+
+"""
+Generate proposal according to θ' = θt + γ(θm − θn) + b
+where γ = 2.38/√(2d) where d is the number of parameters
+* `de`: differential evolution object
+* `Pt`: current particle
+* `group`: a group of particles
+"""
+function variable_gamma(de, Pt, group)
+    Np = length(Pt.Θ)
+    # select the base particle θb
+    Pb = select_base(group)
+    # group without Pb
+    group_diff = setdiff(group, [Pt])
+    # sample particles for θm and θn
+    Pm,Pn = sample(group_diff, 2, replace=false)
+    γ = 2.38 / sqrt(2 * Np)
+    # sample noise for each parameter
+    b = Uniform(-de.ϵ, de.ϵ)
+    # compute proposal value
+    Θp = Pt + γ * (Pm - Pn) + b
+    # reset each parameter to previous value with probability (1-κ)
+    recombination!(de, Pt, Θp)
+    return Θp
+end
+
 """
 Selects base particle θb with probability proportional to weight
 * `group`: a group of particles
 """
-function select_base(group)
+    function select_base(group)
     w = map(x -> x.weight, group)
     θ = exp.(w) / sum(exp.(w))
     p = sample(group, Weights(θ))
-    return p
+return p
 end
 
 """
@@ -72,7 +122,7 @@ function recombination!(de, pt, pp)
             rand() <= (1 - de.κ) ? pp.Θ[i] = pt.Θ[i] : nothing
         end
     end
-    return nothing
+return nothing
 end
 
 # Handles elements within an array
@@ -81,5 +131,5 @@ function recombination!(de, pt, pp, idx)
     for i in 1:N
         rand() <= (1 - de.κ) ? pp.Θ[idx][i] = pt.Θ[idx][i] : nothing
     end
-    return nothing
+return nothing
 end

src/structs.jl

@@ -11,6 +11,8 @@ Differential Evolution MCMC object
 * `κ`: recombination with probability (1-κ) during crossover. Default = 1.0
 * `bounds`: a vector of tuples for lower and upper bounds of each parameter
 * `iter`: current iteration
+* `generate_proposal`: a function that generates proposals. Default is the two mode proposal described in
+Turner et al. 2012. You can also choose fixed_gamma, variable_gamma (see help) or pass a custom function
 
 Constructor signature:
 
@@ -26,7 +28,7 @@ References:
 
 * Turner, B. M., & Sederberg, P. B. (2012). Approximate Bayesian computation with differential evolution. Journal of Mathematical Psychology, 56(5), 375-385.
 """
-mutable struct DE{T1} <: AbstractSampler
+mutable struct DE{T1,F <: Function} <: AbstractSampler
     n_groups::Int64
     Np::Int64
     burnin::Int64
@@ -37,15 +39,16 @@ mutable struct DE{T1} <: AbstractSampler
     κ::Float64
     bounds::T1
     iter::Int64
+    generate_proposal::F
 end
 
 function DE(;n_groups=4, priors=nothing, Np=num_parms(priors) * 3, burnin=1000, α=.1, β=.1, ϵ=.001,
-    σ=.05, κ=1.0, bounds)
+    σ=.05, κ=1.0, bounds, generate_proposal=random_gamma)
     if (α > 0) && (n_groups == 1)
         α = 0.0
         @warn "migration probability α > 0 but n_groups == 1. Changing α = 0.0"
     end
-    return DE(n_groups, Np, burnin, α, β, ϵ, σ, κ, bounds, 1)
+    return DE(n_groups, Np, burnin, α, β, ϵ, σ, κ, bounds, 1, generate_proposal)
 end
 
 """