Skip to content

Pharmacy service access

MaartenHilferink edited this page Sep 10, 2026 · 1 revision

Pharmacy service access — the ServiceAccess branch

Since September 2026. This page describes the components of the ServiceAccess branch and how data flows through them. The method is on Lambda sweep method, the numbers on Pharmacy results September 2026. The school-era scripts that the older Service allocation procedure page describes (greedy-nn.jl, lp-greedy.jl, mip.jl) no longer exist; their formulation survives as run_scenario in lp_run.jl, driven by lp.jl and greedy.jl.

Data flow

 GeoDMS  (cfg/main.dms, STUDY_AREA=<area>)                       per study area
 ├─ NetworkSetup       TomTom road network → cleaned, per-landbody connected network
 ├─ Analyses/…/Allocation  impedance_matrix_od64: every populated 1 km² cell to every
 │                      candidate cell within its choice set  → <area>_od.arrow
 │                                                              <area>_i.arrow  (clients)
 │                                                              <area>_j.arrow  (candidates)
 └─ Analyses/Pharmacies/Descriptives   catchment statistics of today's pharmacies
                          │
                          ▼
 Julia  (lambda_sweep_simplex.jl → lp_run.jl → settings.jl)
 ├─ baseline            every client → nearest existing pharmacy by road
 ├─ λ-sweep             one LP per λ on the same model, warm-started dual simplex (HiGHS)
 ├─ rounding            LP fractions → a real set of p pharmacies (multistart + swaps)
 └─ S1 / S2 / S3        equal count · equal travel · the whole frontier
                          │  logs/sweep_<area>_<FUNC>.log
                          │  <area>/lambda_sweep/<FUNC>/<w=…|S1|S2>/{assignment,traveltime}.arrow
                          ▼
 doc/  (Python + Node)   parse logs → charts → per-region slides → deck  ·  frontier metrics
 GeoDMS                  reads the S1/S2 open sets back (SweepResults) for mapping

Two travel-cost functions are swept in parallel — LINEAR c(t) = t (minutes) and an adapted-logit LOGISTIC c(t) = 1 / (1 + e^{−(t−25)/10}) — and every result exists for both. Two sides are computed per area: the observed network (ExistingPharmacies, the baseline) and the candidate set (NewPharmacies: every cell with ≥ 50 residents, plus every cell that holds a pharmacy today).

The Arrow files

All under C:\LocalData\networkmodel_eu\{ExistingPharmacies,NewPharmacies}\:

File Columns Notes
<area>_od.arrow client_rel, facility_rel, t_ij, d_ij, … one row per OD pair the road network delivers within the choice set; t_ij in seconds
<area>_i.arrow id, pop, total_pop, x, y, NUTS the clients — the same full-population list on both sides, so client_rel is comparable across them
<area>_j.arrow id, x, y the facilities; id is a per-file local index, so existing and candidate facilities can only be matched by (x, y)
<area>/lambda_sweep/<FUNC>/<label>/assignment.arrow id, open one folder per sweep point (w=…), plus S1, S2, and baseline on the existing side
<area>/lambda_sweep/<FUNC>/<label>/traveltime.arrow id, t_ij per client, minutes to its assigned open facility
`_<s1 s2>_<linear logistic>_open.arrow`

GeoDMS (cfg/)

Item Role
cfg/main.dms, ModelParameters.dms Study area from the STUDY_AREA environment variable (Expand(., '%env:STUDY_AREA%')), grid size, Client := 'population', the choice-set knobs max_nr_facilities_per_client := 5 and max_traveltime_to_facility := 120 min.
NetworkSetup.dms, Templates.dms Builds the working road network from the TomTom extract (CreateInitialWorkingNetwork → CreateMoreEfficientNetwork). Keeps the largest strongly-connected component per landbody, not one per study area, so islands and their ferry links survive.
Analyses.dms → Allocate_T The OD matrix: impedance_matrix_od64('bidirectional(link_flag);startPoint(Node_rel);endPoint(Node_rel);cut(OrgZone_max_imp);limit(OrgZone_max_mass,DstZone_mass);alternative(link_imp):alt_imp;od:OrgZone_rel,DstZone_rel,impedance', …). cut = 120 min; limit = 5 with mass 1 only on existing-pharmacy cells, so a client's choice set is every candidate within the road time to its 5th-nearest existing pharmacy. Exports DistanceTableExport, ClientExport, FacilityExport.
Analyses/AllocateClientsToExistingPharmacies, …ToNewPharmacies The two sides, each an Allocate_T instance.
Analyses/Pharmacies/Descriptives Today's network by road catchment: residents per pharmacy and per 1 km² location, empty (never-nearest) pharmacies, catchment percentiles. One Arrow row per area, gathered by collect_descriptives.jl.
Analyses/…/SweepResults (ReadSweepResults_T) Reads baseline, S1, S2 per travel function back into GeoDMS.
SourceData/Locations.dms services_allocated Reads the exported open-location files.

Julia (repository root)

File Role
settings.jl Paths, COUNTRIES, the travel-cost functions c(t), big_cost(), facility-cost placeholders (FACILITY_MIN_COSTS = 100 000), Arrow loaders, the candidate-subsampling knob, and the region-exclusion rule. Everything is env-overridable.
lp_run.jl The engine. build_lp_warmstart / solve_at_w! — the soft-coverage LP, solved along the λ-grid from the previous basis; greedy_round, multistart_round, swap_round!, travel_of, assign_nearest — rounding and scoring; run_scenario — the older hard-coverage LP with a minimum-catchment penalty.
lambda_sweep_simplex.jl The sweep driver used for all results. Baseline metrics, the common w-grid, per-w LP → rounding → log line, S1/S2 bracketing, the per-point Arrow outputs.
lambda_sweep.jl Earlier parallel variant (one fresh LP per w). Superseded; kept for reference.
lp.jl, greedy.jl, greedy-merge.jl School-era scenario runners over run_scenario: hard coverage, min_clients deficit penalty, IPM; grid searches. Not used for the pharmacy results.
cap_scenario.jl, s1_cap.jl, s2_cap.jl The catchment-cap ladder without a cost function (S1-A/B/C, S2-A/B). Explored on the Netherlands; not pursued.
export_s1s2_locations.jl Joins each area's S1/S2 assignment.arrow to <area>_j.arrow and writes the open locations.
collect_descriptives.jl Gathers the descriptive Arrow rows into doc/pharmacy_descriptives.csv (+ _nuts1.csv).

Orchestration

Script Role
run_pharmacy_pipeline.bat <area…> GeoDMS, existing side: steps network1 network2 alloc (subset via STEPS).
run_new_pharmacy_pipeline.bat <area…> GeoDMS, candidate side: the same three steps.
run_new_country_sweeps.ps1 Orchestrator for a new area: both pipelines, then the LINEAR and LOGISTIC sweeps; resumable; progress to logs/orchestrator_progress.log.
run_rebuild_all.ps1, run_realloc_all.ps1 Rebuild networks + ODs on both sides / re-export only the client and facility tables.
run_recalc_batch.ps1, run_resweep_batch.ps1 Per-area workers: rebuild + re-sweep, or sweep only. Several run side by side in visible consoles.
run_cap_scenarios.bat, make_descriptive_table.bat The cap-ladder runs; the descriptive table over all areas.

Deck and metrics (doc/)

build_deck_data.py (logs → deck_data.json) → build_charts.py (per-region charts) → build_deck.mjs (region + summary slides, pptxgenjs) → merge_deck.ps1 (concept slides + generated slides → lambda_sweep5.pptx). frontier_metrics.py computes the improvement rectangle, the diagonal crossing and λ at the crossing per area (frontier_metrics_interp3.csv); frontier_charts.py draws the ranked lists; policy_typology.py attaches the regulation typology used to colour them. doc/README_deck.md has the run order.

Running it

Prerequisites: GeoDMS 20.19.1 or later (GEODMS_EXE), the TomTom extract under %NetworkModelDataDir%\Infrastructure\TomTom\<area>\, GeoDMS's LocalDataDir/SourceDataDir settings, and Julia with Arrow, JuMP, HiGHS (Installation of Julia).

run_pharmacy_pipeline.bat Netherlands          :: existing side
run_new_pharmacy_pipeline.bat Netherlands      :: candidate side
set TRAVEL_FUNC=LINEAR   && julia lambda_sweep_simplex.jl   > logs\sweep_Netherlands_LINEAR.log
set TRAVEL_FUNC=LOGISTIC && julia lambda_sweep_simplex.jl   > logs\sweep_Netherlands_LOGISTIC.log

or run_new_country_sweeps.ps1 for a new area end to end. Sweep knobs are environment variables — COUNTRIES, TRAVEL_FUNC, SWEEP_WMAX, SWEEP_MULTS, SOLVER, LP_TIME_LIMIT, LOCATION_SELECTION_FACTOR, PROTECT_BASELINE, ROUNDING, MS_RESTARTS/ROUNDS/SEED, BIG_TRAVELTIME_MIN, LOGISTIC_MIDPOINT/SCALE, NUTS_EXCLUSION, NUTS_EXCLUDE_SHARE — listed with their defaults in settings.jl.

Sweep cost: minutes per λ-point for most areas, hours for the largest (Poland country-level: 15.1 M OD rows, 29 h LINEAR). The largest areas are candidate-subsampled (LOCATION_SELECTION_FACTOR), keeping every baseline location so the frontier still dominates the baseline.

Clone this wiki locally