Switch from Euclidean to approximate driving distance

src/RELOG.jl

@@ -5,19 +5,19 @@
 module RELOG
 
 include("instance/structs.jl")
-
 include("graph/structs.jl")
 
+include("instance/geodb.jl")
+include("graph/dist.jl")
 include("graph/build.jl")
 include("graph/csv.jl")
 include("instance/compress.jl")
-include("instance/geodb.jl")
 include("instance/parse.jl")
 include("instance/validate.jl")
 include("model/build.jl")
 include("model/getsol.jl")
-include("model/solve.jl")
 include("model/resolve.jl")
+include("model/solve.jl")
 include("reports/plant_emissions.jl")
 include("reports/plant_outputs.jl")
 include("reports/plants.jl")

src/graph/build.jl

@@ -2,14 +2,6 @@
 # Copyright (C) 2020, UChicago Argonne, LLC. All rights reserved.
 # Released under the modified BSD license. See COPYING.md for more details.
 
-using Geodesy
-
-function calculate_distance(source_lat, source_lon, dest_lat, dest_lon)::Float64
-    x = LLA(source_lat, source_lon, 0.0)
-    y = LLA(dest_lat, dest_lon, 0.0)
-    return round(euclidean_distance(x, y) / 1000.0, digits = 2)
-end
-
 function build_graph(instance::Instance)::Graph
     arcs = []
     next_index = 0
@@ -48,13 +40,15 @@ function build_graph(instance::Instance)::Graph
     end
 
     # Build arcs from collection centers to plants, and from one plant to another
+    metric = _KnnDrivingDistance()
     for source in [collection_shipping_nodes; plant_shipping_nodes]
         for dest in process_nodes_by_input_product[source.product]
-            distance = calculate_distance(
+            distance = _calculate_distance(
                 source.location.latitude,
                 source.location.longitude,
                 dest.location.latitude,
                 dest.location.longitude,
+                metric,
             )
             values = Dict("distance" => distance)
             arc = Arc(source, dest, values)

src/graph/dist.jl

@@ -0,0 +1,69 @@
+# RELOG: Reverse Logistics Optimization
+# Copyright (C) 2020, UChicago Argonne, LLC. All rights reserved.
+# Released under the modified BSD license. See COPYING.md for more details.
+
+using Geodesy
+using NearestNeighbors
+using DataFrames
+
+Base.@kwdef mutable struct _KnnDrivingDistance
+    tree = nothing
+    ratios = nothing
+end
+
+mutable struct _EuclideanDistance end
+
+function _calculate_distance(
+    source_lat,
+    source_lon,
+    dest_lat,
+    dest_lon,
+    ::_EuclideanDistance,
+)::Float64
+    x = LLA(source_lat, source_lon, 0.0)
+    y = LLA(dest_lat, dest_lon, 0.0)
+    return round(euclidean_distance(x, y) / 1000.0, digits = 3)
+end
+
+function _calculate_distance(
+    source_lat,
+    source_lon,
+    dest_lat,
+    dest_lon,
+    metric::_KnnDrivingDistance,
+)::Float64
+    if metric.tree === nothing
+        basedir = joinpath(dirname(@__FILE__), "..", "..", "data")
+        csv_filename = joinpath(basedir, "dist_driving.csv")
+
+        # Download pre-computed driving data
+        if !isfile(csv_filename)
+            _download_zip(
+                "https://axavier.org/RELOG/0.6/data/dist_driving_0b9a6ad6.zip",
+                basedir,
+                csv_filename,
+                0x0b9a6ad6,
+            )
+        end
+
+        # Fit kNN model
+        df = DataFrame(CSV.File(csv_filename))
+        coords = Matrix(df[!, [:source_lat, :source_lon, :dest_lat, :dest_lon]])'
+        metric.ratios = Matrix(df[!, [:ratio]])
+        metric.tree = KDTree(coords)
+    end
+
+    # Compute Euclidean distance
+    dist_euclidean = _calculate_distance(
+        source_lat,
+        source_lon,
+        dest_lat,
+        dest_lon,
+        _EuclideanDistance(),
+    )
+
+    # Predict ratio
+    idxs, _ = knn(metric.tree, [source_lat, source_lon, dest_lat, dest_lon], 5)
+    ratio_pred = mean(metric.ratios[idxs])
+    return round(dist_euclidean * ratio_pred, digits = 3)
+end