From 79fb4eb34e80736561419efb465870043c715656 Mon Sep 17 00:00:00 2001
From: dcradu <dcradu@uliege.be>
Date: Sun, 18 Apr 2021 21:54:47 +0200
Subject: [PATCH] updated SA algorithms (time gains)
---
src/jl/MCP_heuristics.jl | 112 +++++++++++++++++++++++++++++++++++++
src/jl/SitingHeuristics.jl | 9 +--
2 files changed, 117 insertions(+), 4 deletions(-)
diff --git a/src/jl/MCP_heuristics.jl b/src/jl/MCP_heuristics.jl
index cfc35ad..6e01c8f 100644
--- a/src/jl/MCP_heuristics.jl
+++ b/src/jl/MCP_heuristics.jl
@@ -485,4 +485,116 @@ function stochastic_threshold_greedy_algorithm_trajectory(D::Array{Float64,2}, c
end
return ind_incumbent, obj_incumbent
+end
+
+#################### Simulated Annealing Local Search #######################
+
+# Description: function implementing a simulated annealing-inspired local search for unpartitioned geographical regions
+#
+# Comments: 1) types of inputs should match those declared in argument list
+#
+# Inputs: D - criticality matrix with entries in {0, 1}, where rows represent time windows and columns represent locations
+# c - global criticality threshold
+# n - number of sites to deploy
+# N - number of locations to swap in order to obtain a neighbour of the incumbent solution
+# I - number of iterations (outer loop), defines the number of times the incumbent solution may be updated
+# E - number of epochs (inner loop), defines the number of neighbours of the incumbent solution sampled at each iteration
+# x_init - initial solution, vector with entries in {0, 1}, with cardinality n and whose dimension is compatible with D
+# T_init - initial temperature from which the (exponentially-decreasing) temperature schedule is constructed
+#
+# Outputs: x_incumbent - vector with entries in {0, 1} and cardinality n representing the incumbent solution at the last iteration
+# LB_incumbent - objective value of incumbent solution, provides a lower bound on optimal objective
+# obj - vector storing the incumbent objective value at each iteration
+#
+
+function simulated_annealing_local_search(D::Array{Float64, 2}, c::Float64, n::Float64, N::Int64, I::Int64, E::Int64, x_init::Array{Float64, 1}, T_init::Float64, legacy_locations::Vector{Int64})
+
+ W, L = size(D)
+ n = convert(Int64, n)
+ n_new = n - length(legacy_locations)
+
+ # Pre-allocate lower bound vector
+ obj = Vector{Int64}(undef, I)
+
+ # Pre-allocate x-related arrays
+ ind_ones_incumbent = Vector{Int64}(undef, n_new)
+ ind_ones_incumbent_filtered = Vector{Int64}(undef, n_new-N)
+ ind_zeros_incumbent = Vector{Int64}(undef, L-n)
+ ind_zeros_incumbent_filtered = Vector{Int64}(undef, L-n-N)
+ ind_ones2zeros_candidate = Vector{Int64}(undef, N)
+ ind_zeros2ones_candidate = Vector{Int64}(undef, N)
+ ind_ones2zeros_tmp = Vector{Int64}(undef, N)
+ ind_zeros2ones_tmp = Vector{Int64}(undef, N)
+
+ # Pre-allocate y-related arrays
+ y_incumbent = Array{Bool}(undef, W, 1)
+ y_tmp = Array{Bool}(undef, W, 1)
+
+ Dx_incumbent = Array{Float64}(undef, W, 1)
+ Dx_tmp = Array{Float64}(undef, W, 1)
+
+ # Initialise
+ ind_deployed = findall(x_init .== 1.)
+ Dx_incumbent .= sum(view(D, :, ind_deployed), dims = 2)
+ y_incumbent .= Dx_incumbent .>= c
+ ind_ones_incumbent .= filter(a -> !(a in legacy_locations), ind_deployed)
+ ind_zeros_incumbent .= findall(x_init .== 0.)
+
+ # Iterate
+ @inbounds for i = 1:I
+ obj[i] = sum(y_incumbent)
+ delta_candidate = -100000
+ @inbounds for e = 1:E
+ # Sample from neighbourhood
+ sample!(ind_ones_incumbent, ind_ones2zeros_tmp, replace=false)
+ sample!(ind_zeros_incumbent, ind_zeros2ones_tmp, replace=false)
+
+ # Compute y and associated objective value
+ @inbounds for j = 1:N
+ Dx_tmp .= Dx_incumbent .+ view(D, :, ind_zeros2ones_tmp[j]) .- view(D, :, ind_ones2zeros_tmp[j])
+ end
+ y_tmp .= Dx_tmp .>= c
+
+ # Update objective difference
+ delta_tmp = sum(y_tmp) - obj[i]
+
+ # Update candidate solution
+ if delta_tmp > delta_candidate
+ ind_ones2zeros_candidate .= ind_ones2zeros_tmp
+ ind_zeros2ones_candidate .= ind_zeros2ones_tmp
+ delta_candidate = delta_tmp
+ end
+ end
+ if delta_candidate > 0
+ ind_ones_incumbent_filtered .= filter(a -> !(a in ind_ones2zeros_candidate), ind_ones_incumbent)
+ ind_ones_incumbent[1:(n_new-N)] .= ind_ones_incumbent_filtered
+ ind_ones_incumbent[(n_new-N+1):n_new] .= ind_zeros2ones_candidate
+ ind_zeros_incumbent_filtered .= filter(a -> !(a in ind_zeros2ones_candidate), ind_zeros_incumbent)
+ ind_zeros_incumbent[1:(L-n-N)] .= ind_zeros_incumbent_filtered
+ ind_zeros_incumbent[(L-n-N+1):(L-n)] .= ind_ones2zeros_candidate
+ @inbounds for j = 1:N
+ Dx_incumbent .= Dx_incumbent .+ view(D, :, ind_zeros2ones_candidate[j]) .- view(D, :, ind_ones2zeros_candidate[j])
+ end
+ y_incumbent .= Dx_incumbent .>= c
+ else
+ T = T_init * exp(-10*i/I)
+ p = exp(delta_candidate / T)
+ d = Binomial(1, p)
+ b = rand(d)
+ if b == 1
+ ind_ones_incumbent_filtered .= filter(a -> !(a in ind_ones2zeros_candidate), ind_ones_incumbent)
+ ind_ones_incumbent[1:(n_new-N)] .= ind_ones_incumbent_filtered
+ ind_ones_incumbent[(n_new-N+1):n_new] .= ind_zeros2ones_candidate
+ ind_zeros_incumbent_filtered .= filter(a -> !(a in ind_zeros2ones_candidate), ind_zeros_incumbent)
+ ind_zeros_incumbent[1:(L-n-N)] .= ind_zeros_incumbent_filtered
+ ind_zeros_incumbent[(L-n-N+1):(L-n)] .= ind_ones2zeros_candidate
+ @inbounds for j = 1:N
+ Dx_incumbent .= Dx_incumbent .+ view(D, :, ind_zeros2ones_candidate[j]) .- view(D, :, ind_ones2zeros_candidate[j])
+ end
+ y_incumbent .= Dx_incumbent .>= c
+ end
+ end
+ end
+ obj_incumbent = sum(y_incumbent)
+ return ind_ones_incumbent, obj_incumbent, obj
end
\ No newline at end of file
diff --git a/src/jl/SitingHeuristics.jl b/src/jl/SitingHeuristics.jl
index b0954cf..8892e69 100644
--- a/src/jl/SitingHeuristics.jl
+++ b/src/jl/SitingHeuristics.jl
@@ -35,11 +35,12 @@ function main_MIRSA(index_dict, deployment_dict, D, c, N, I, E, T_init, R, run,
run = string(run)
p = string(p)
data_path = string(data_path)
+ legacy_index = Vector{Float64}(undef, 0)
W, L = size(D)
P = maximum(values(index_dict))
- n_partitions = [deployment_dict[i] for i in 1:P]
+ n = deployment_dict[1]
if run == "MIR"
@@ -47,7 +48,7 @@ function main_MIRSA(index_dict, deployment_dict, D, c, N, I, E, T_init, R, run,
x_init = solve_MILP_partitioning(D, c, n_partitions, index_dict, "Gurobi")
for r = 1:R
- x_sol[r, :], LB_sol[r], obj_sol[r, :] = simulated_annealing_local_search_partition(D, c, n_partitions, N, I, E, x_init, T_init, index_dict)
+ x_sol[r, :], LB_sol[r], obj_sol[r, :] = simulated_annealing_local_search(D, c, n, N, I, E, x_init, T_init, legacy_index)
end
elseif run == "RS"
@@ -68,7 +69,7 @@ function main_MIRSA(index_dict, deployment_dict, D, c, N, I, E, T_init, R, run,
for r = 1:R
println(r)
- x_sol[r, :], LB_sol[r], obj_sol[r, :] = simulated_annealing_local_search_partition(D, c, n_partitions, N, I, E, x_init, T_init, index_dict)
+ x_sol[r, :], LB_sol[r], obj_sol[r, :] = simulated_annealing_local_search(D, c, n, N, I, E, x_init, T_init, legacy_index)
end
elseif run == "SGH"
@@ -85,7 +86,7 @@ function main_MIRSA(index_dict, deployment_dict, D, c, N, I, E, T_init, R, run,
for r = 1:R
println(r)
- x_sol[r, :], LB_sol[r], obj_sol[r, :] = simulated_annealing_local_search_partition(D, c, n_partitions, N, I, E, x_init_best, T_init, index_dict)
+ x_sol[r, :], LB_sol[r], obj_sol[r, :] = simulated_annealing_local_search(D, c, n, N, I, E, x_init_best, T_init, legacy_index)
end
else
--
GitLab