EMS Readiness Optimization for Manhattan

A simulation-based approach to optimizing ambulance staging locations in Manhattan, reducing EMS response times by 19% over a spatially-stratified baseline through demand-weighted allocation with capacity=2 constraints.

Key Findings

Metric	P0 (Baseline)	P2 (Optimized)	Improvement
Mean Response Time	3.17 min	2.57 min	−18.9%
P95 Response Time	6.26 min	4.66 min	−25.6%
8-min Coverage	99.7%	99.7%	+0.0 pp
Mean Utilization	7.6%	7.4%	−0.2 pp

Results based on 2,700+ simulation experiments with 30 replications each (p < 0.001). Capacity sensitivity analysis (cap 1–5) confirms cap=2 as operationally optimal default. Alternative analyses confirm stability under distance metric choice and geographic focus.

Project Overview

This project evaluates three ambulance allocation policies for Manhattan's 48 FDNY firehouses:

P0 (Spatially-Stratified Uniform): Latitude-based geographic stratification providing even north–south coverage
P1 (Demand-Proportional): Units allocated based on nearby demand
P2 (Demand-Weighted Optimized): MIP-optimized allocation minimizing expected response time

The analysis combines:

NHPP Demand Modeling — Calibrated from 2.24M historical motor vehicle collision records
Mixed-Integer Programming — Three optimization formulations with capacity=2 constraint (PuLP/CBC solver)
Discrete-Event Simulation — SimPy-based DES with 2,700+ production runs across 9 experiment sets
Statistical Analysis — ANOVA, Tukey HSD, confidence intervals, effect sizes
Capacity Sensitivity Analysis — Full spectrum (cap 1–5) across 9 fleet sizes
Alternative Analyses — Manhattan distance metric comparison + CBD-focused optimization evaluation

Start Here

If you're new to the project:

Read the Technical Report §1 (Executive Summary)
Browse results/baseline/ — this is the canonical source of truth
Run the baseline pipeline (optional — results are pre-committed):
```
python scripts/run_production_v2.py
```
Understand the tier system: Canonical → Supporting → Historical (see below)

For supporting analyses: results/analysis/ and docs/analysis/
For architecture deep-dive: docs/core/ARCHITECTURAL_MAP.md
For artifact governance: docs/core/artifact_governance_map.md

Source of Truth

All results and documentation follow a three-tier governance system:

Tier	Label	Location	Status
Canonical	✅ Current	`results/baseline/`, `docs/core/`	Authoritative — cited in technical report and README
Supporting	🔬 Analysis	`results/analysis/`, `docs/analysis/`	Robustness and sensitivity analyses that strengthen canonical findings
Historical	⚠️ Archive	`results/archive/`, `docs/archive/`, `scripts/archive/`	Superseded artifacts (e.g., cap=5 era). Retained for audit trail only

Canonical values used throughout this project:

Parameter	Value	Source
Base arrival rate (λ₀)	3.48 crashes/hour	`configs/demand.yaml`
Firehouse capacity (C)	2 units/firehouse	`configs/optimization.yaml`, DEC-010
Simulation horizon	168 hours (1 week)	`configs/simulation.yaml`
Replications per scenario	30	`configs/simulation.yaml`
Master seed	42	`configs/demand.yaml`
Coverage threshold	8 minutes (NFPA) / 6 minutes (NYC)	`configs/simulation.yaml`
Average EMS speed	20 mph	`configs/service.yaml`
Service time mean	25 min (LogNormal)	`configs/service.yaml`
Canonical K set	{10, 15, 20, 25, 30, 35, 40, 45, 48}	`configs/optimization.yaml`

Full governance details: docs/core/artifact_governance_map.md Figure/table provenance: docs/core/figure_table_traceability.md

Experimental Scope

The study comprises 4 core experiments and 4 supporting analyses, totalling 2,700+ simulation runs:

Experiment	Factors	Scope	Generator
Exp1 — Policy Comparison	P0, P1, P2 at K=20	3 × 30 reps = 90 runs	`scripts/run_production_v2.py`
Exp2 — Fleet Sensitivity	9 K-values × 3 policies	27 × 30 reps = 810 runs	`scripts/run_production_v2.py`
Exp3 — Demand Sensitivity	6 demand multipliers × 2 policies	12 × 30 reps = 360 runs	`scripts/run_production_v2.py`
Exp4 — Service Robustness	3 service means × 2 policies	6 × 30 reps = 180 runs	`scripts/run_production_v2.py`
Capacity — Cap Sensitivity	cap {1–5} × K {20,40} × 3 policies	Supporting	`scripts/analysis/capacity_sensitivity_*.py`
CBD — CBD Robustness	CBD surge scenarios	330 runs	`scripts/analysis/run_cbd_experiment.py`
Distance — Metric Comparison	Haversine vs Manhattan	Supporting	`scripts/analysis/run_distance_comparison_experiment.py`
Queue/Seasonal — Derived	Post-hoc analysis of Exp1–4	Supporting	`scripts/analysis/analyze_queue_metrics.py`, `analyze_seasonal_patterns.py`

All experiments use Common Random Numbers (CRN) for variance reduction. Verification (4 tests) and validation (3 pilots) are documented in the technical report §4.4.3.

Full experimental design: docs/core/experimental_design.md

Repository Architecture

The repository is organized to eliminate ambiguity:

Results:

Baseline → canonical results (results/baseline/)
Analysis → robustness & sensitivity (results/analysis/)
Archive → historical artifacts (results/archive/)

Scripts:

Scripts (root) → production entry points
scripts/analysis/ → supporting analyses
scripts/archive/ → legacy utilities

Documentation:

docs/core/ → primary documentation
docs/analysis/ → supporting analyses
docs/archive/ → historical records

Repository Structure

ems-optimization/
├── configs/ # YAML configuration files
│ ├── demand.yaml # NHPP demand model parameters
│ ├── optimization.yaml # MIP solver settings (capacity=2 default)
│ ├── service.yaml # Travel & service time config
│ ├── simulation.yaml # DES engine settings
│ └── cbd_scenario.yaml # CBD robustness experiment config
├── data/
│ ├── raw/ # Original data files (NYC Open Data)
│ ├── processed/ # Generated data (not in Git — run `make data`)
│ └── manifests/ # Data audit records
├── docs/ # Project documentation (62+ .md files, see docs/core/DOCUMENTATION_INDEX.md)
│ ├── core/ # Key deliverables, specs, guides
│ │ ├── technical_report.md # Full final report (v2.1.0)
│ │ ├── executive_presentation.md # Stakeholder slide deck
│ │ └── ... # Additional core docs
│ ├── analysis/ # Robustness & sensitivity analyses
│ │ ├── capacity_sensitivity_analysis.md
│ │ └── ... # Additional analysis docs
│ └── archive/ # Historical/superseded docs
├── notebooks/ # Jupyter analysis notebooks
│ ├── 02_eda_spatiotemporal.ipynb
│ ├── 03_input_modeling.ipynb
│ ├── 04_service_travel_proxy.ipynb
│ ├── 05_optimization.ipynb
│ ├── 06_simulation_debug.ipynb
│ ├── 07_production_results.ipynb
│ ├── 08_statistical_analysis.ipynb
│ ├── 09_cbd_analysis.ipynb
│ └── colab_standalone/ # Google Colab standalone notebooks
│     ├── individual/ # 8 phase notebooks (00–07)
│     ├── EMS_Optimization_Complete_Pipeline.ipynb
│     └── README.md
├── results/
│ ├── baseline/ # Canonical production results (cap=2)
│ │ ├── allocations/ # Policy allocation CSVs
│ │ ├── figures/ # Publication-ready figures
│ │ ├── simulation/ # Verification & validation output
│ │ └── tables/ # Statistical summary tables
│ ├── analysis/ # Supporting robustness & sensitivity
│ │ ├── capacity_comparison/ # Capacity sensitivity (cap 1–5) results
│ │ ├── cbd_focused_comparison/ # CBD-focused vs Manhattan-wide results
│ │ ├── distance_comparison/ # Haversine vs Manhattan metric results
│ │ ├── heatmaps/ # Allocation heatmaps (K × cap × policy)
│ │ ├── maps/ # Allocation map visualizations
│ │ ├── simulation/ # CBD & production DES output
│ │ └── tables/ # Analysis summary tables
│ ├── archive/ # Legacy/historical artifacts (cap=5 era)
│ │ ├── figures/ # Superseded figures
│ │ ├── optimization/ # Phase 3 historical optimization results
│ │ └── tables/ # Superseded tables
│ ├── figures/ # Top-level visualization PNGs
│ ├── tables/ # Top-level statistical tables (CSV + LaTeX)
│ └── simulation/ # Top-level simulation output data
├── scripts/ # Automation & analysis scripts (23+)
├── src/ems_readiness/ # Core Python package (v0.6.0)
│ ├── demand/ # NHPP arrival generator
│ ├── service/ # Travel time & service time models
│ ├── optimization/ # MIP formulations, policies & allocator
│ ├── simulation/ # SimPy DES engine
│ └── utils/ # Distance calculations
├── tests/ # 176 unit tests (pytest)
├── requirements.txt # Python dependencies
├── Makefile # Build automation
└── LICENSE # MIT License

Installation & Setup

Prerequisites

Python 3.11+
pip

Quick Start

# Clone the repository
git clone https://github.com/cnsp/ems-optimization.git
cd ems-optimization

# Create virtual environment and install dependencies
make setup
# Or manually:
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# Install project package in development mode
pip install -e .

Data Setup

Large raw data files are tracked via Git LFS. Ensure Git LFS is installed before cloning:

git lfs install   # one-time setup
git clone https://github.com/cnsp/ems-optimization.git

The following files are pulled automatically via LFS on clone:

data/raw/Motor_Vehicle_Collisions_-_Crashes_20260223.csv (536 MB)
data/raw/Police_Precincts_20260223.csv (3.6 MB)

All other raw files (boundary GeoJSONs, firehouse listings, data dictionaries) and key processed files (distance matrices, demand lambdas, firehouse lists) are tracked directly in Git.

If LFS files were not pulled (e.g., shallow clone), run:

git lfs pull

See data/raw/README.md for original NYC Open Data download links.

Note: The Colab notebooks (notebooks/colab_standalone/) include self-contained data processing cells that generate crashes_manhattan.csv from the raw crash CSV on first run. All other processed files are tracked in Git and require no extra setup.

Generate all processed data (single command):

make data
# Or: python scripts/generate_all_data.py

Verify data:

make verify-data

Pipeline CLI Options

The data pipeline (scripts/generate_all_data.py) supports several options:

Flag	Description
`--force`	Regenerate all files even if they exist
`--no-validate`	Skip raw data validation checks
`--no-cache`	Disable smart caching
`--jobs N` / `-j N`	Parallel workers (0 = auto-detect CPU count)
`--tier {1,2,3}`	Run only a specific tier
`--version`	Show current data version info
`--seed N`	Master seed for reproducibility (default: 42)
`--verify`	Check data existence without generating

# Force-regenerate with 4 parallel workers
python scripts/generate_all_data.py --force --jobs 4

# Show data version/lineage info
python scripts/generate_all_data.py --version

# Skip validation (faster startup)
python scripts/generate_all_data.py --no-validate

Note: Notebooks auto-detect and regenerate missing processed data on first run, so you can also just open any notebook and it will trigger the pipeline if needed.

How to Reproduce Results

Run the Full Pipeline

# 1. Process data (if raw data available)
make data

# 2. Run optimization comparison
python scripts/analysis/run_optimization_comparison.py

# 3. Run simulation verification
python scripts/run_verification.py

# 4. Run validation pilots
python scripts/run_validation_pilots.py

# 5. Run production experiments (1,440 simulations — takes ~2-4 hours)
python scripts/run_production_v2.py

# 6. Analyze results and generate figures
python scripts/analysis/analyze_production_results.py
python scripts/analysis/generate_publication_figures.py

# 7. Run CBD robustness experiment (330 simulations)
python scripts/analysis/run_cbd_experiment.py --reps 30

# 8. Run gap closure analyses
python scripts/analysis/analyze_queue_metrics.py
python scripts/analysis/analyze_seasonal_patterns.py

# 9. Run alternative analyses (Phase 8)
python scripts/analysis/generate_manhattan_distance_matrix.py
python scripts/analysis/run_distance_comparison_experiment.py --reps 10
python scripts/analysis/run_cbd_focused_optimization.py --reps 10

# 10. Run capacity sensitivity analysis (Phase 9)
python scripts/analysis/capacity_sensitivity_full_spectrum.py

# 11. Run P0 spatial stratification analysis
python scripts/analysis/p0_spatial_analysis.py

# 12. Run extended fleet analysis experiments (810 simulations — cap=2)
python scripts/run_production_v2.py

Run Specific Experiments

from ems_readiness.simulation import EMSSimulation, BatchRunner
from ems_readiness.optimization import EMSAllocator

# Initialize allocator
allocator = EMSAllocator.from_project(".")

# Get P2 allocation for K=20 units (capacity=2 default)
result = allocator.solve(model="demand_weighted", K=20)
print(result.allocation)

# Run simulation
sim = EMSSimulation.from_config("configs/simulation.yaml", allocation=result.allocation)
metrics = sim.run(duration_hours=168)
print(f"Mean RT: {metrics['mean_response_time']:.2f} min")

How to Run Tests

# Run all 176 tests
pytest tests/ -v

# Run specific test modules
pytest tests/test_simulation_core.py -v
pytest tests/test_dispatch_logic.py -v
pytest tests/test_extreme_cases.py -v
pytest tests/test_reproducibility.py -v

Documentation Index

📖 Full index: docs/core/DOCUMENTATION_INDEX.md — master list of all 62 docs with status badges (✅ Current / 🔄 Historical / 📋 Reference)

Key Documents

Document	Description
Technical Report	Full final report (v2.1.0) with all findings
Executive Summary	One-page executive summary
Executive Presentation	10-slide stakeholder presentation
Data Usage Guide	Which data/results files to use and why
Reproducibility Guide	How to reproduce all results

Model & Methods

Document	Description
Conceptual Model	DES model specification
Optimization Formulation	MIP mathematical formulations
Experimental Design	Factorial experiment specification
Optimization Results	Current optimization results (cap=2)

Analysis & Robustness

Document	Description
Capacity Sensitivity	Full-spectrum capacity (cap 1–5) analysis
CBD Robustness	CBD-specific DES experiment findings
Distance Metric Comparison	Haversine vs. Manhattan distance analysis
Research Questions	Evaluation of all 5 research questions

Guides

Document	Description
Code Documentation	Architecture and API guide
Notebook Guide	Notebook descriptions and execution order
Testing Guide	Testing framework and how to run tests
Figure Trace Guide	Figure-to-source traceability
Visualization Index	Complete catalog of all generated figures

Navigating Results

The results/ folder contains output from multiple project phases. Each subfolder has its own README.md explaining what's there and whether it's current or legacy.

Folder	Label	Status	What's Inside
`results/baseline/`	Canonical	✅ Current	Production allocations, figures, simulation verification/validation, tables
`results/baseline/simulation/`	Canonical	✅ Current	Verification tests, validation pilots
`results/analysis/`	Supporting	✅ Current	All robustness & sensitivity analyses
`results/analysis/capacity_comparison/`	Supporting	✅ Current	Full capacity sweep (cap 1–5)
`results/analysis/cbd_focused_comparison/`	Supporting	✅ Current	CBD-focused optimization experiment
`results/analysis/distance_comparison/`	Supporting	✅ Current	Manhattan vs Haversine comparison
`results/analysis/heatmaps/`	Supporting	✅ Current	Allocation heatmaps across K × cap × policy
`results/analysis/maps/`	Supporting	✅ Current	Allocation maps at K=40
`results/archive/`	Legacy	⚠️ Historical	Phase 3 historical artifacts (cap=5 era)
`results/archive/optimization/`	Legacy	⚠️ Historical	Superseded optimization results
`results/figures/`	—	✅ Mostly Current	Publication figures (`pub_fig*`) + technical plots
`results/tables/`	—	✅ Mostly Current	Publication tables + experiment pivots (see README)
`results/simulation/`	—	✅ Current	Top-level simulation output data

Quick reference: results/WHICH_FILES_TO_USE.md
Full guide: docs/core/data_usage_guide.md

Google Colab Notebooks

For external verification and reproduction, standalone Google Colab notebooks are available in notebooks/colab_standalone/. These require no local setup — just open in Colab and run.

Notebook	Phase	Runtime
`00_colab_setup_and_data`	Environment setup, data verification	~2 min
`01_colab_eda_spatiotemporal`	Exploratory data analysis	~5 min
`02_colab_demand_modeling`	NHPP demand model	~5 min
`03_colab_service_modeling`	Travel/service time models	~3 min
`04_colab_optimization`	P0/P1/P2 allocation policies	~5 min
`05_colab_simulation`	SimPy DES (demo or full-scale)	5 min – 8 hr
`06_colab_statistical_analysis`	ANOVA, effect sizes, CIs	~3 min
`07_colab_visualization_reporting`	Publication figures, summary report	~3 min
All-in-One Pipeline	Complete end-to-end analysis	5–10 hr

See notebooks/colab_standalone/README.md for detailed instructions.

Research Questions

RQ1: How does demand for EMS services vary spatially and temporally across Manhattan?
RQ2: What is the optimal allocation of K ambulances to minimize expected response time?
RQ3: How do optimized allocations compare to the spatially-stratified uniform baseline under realistic conditions?
RQ4: How sensitive are policy rankings to fleet size, demand intensity, service time, and firehouse capacity?
RQ5: What fleet size achieves ≥95% coverage within 8 minutes?

Citation

@misc{ems_optimization_2026,
 title={EMS Readiness Optimization for Manhattan: A Simulation-Based Approach to Ambulance Staging},
 author={EMS Optimization Research Team},
 year={2026},
 howpublished={\url{https://github.com/cnsp/ems-optimization}},
 note={Version 1.3.0}
}

License

This project is licensed under the MIT License — see LICENSE for details.

Contact

Repository: github.com/cnsp/ems-optimization
Issues: GitHub Issues

Built with Python, SimPy, PuLP, pandas, and matplotlib. 8,500+ lines of code across 14 modules, tested with 176 unit tests and 2,700+ simulation experiments. Covers capacity sensitivity (cap 1–5), spatially-stratified P0 baseline, and alternative distance metric and CBD-focused optimization analyses.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

EMS Readiness Optimization for Manhattan

Key Findings

Project Overview

Start Here

Source of Truth

Experimental Scope

Repository Architecture

Repository Structure

Installation & Setup

Prerequisites

Quick Start

Data Setup

Pipeline CLI Options

How to Reproduce Results

Run the Full Pipeline

Run Specific Experiments

How to Run Tests

Documentation Index

Key Documents

Model & Methods

Analysis & Robustness

Guides

Navigating Results

Google Colab Notebooks

Research Questions

Citation

License

Contact

About

Releases

Packages

Used by

Contributors

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 104 Commits
configs		configs
data		data
docs		docs
notebooks		notebooks
results		results
scripts		scripts
src/ems_readiness		src/ems_readiness
tests		tests
.abacus.donotdelete		.abacus.donotdelete
.gitattributes		.gitattributes
.gitignore		.gitignore
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation

EMS Readiness Optimization for Manhattan

Key Findings

Project Overview

Start Here

Source of Truth

Experimental Scope

Repository Architecture

Repository Structure

Installation & Setup

Prerequisites

Quick Start

Data Setup

Pipeline CLI Options

How to Reproduce Results

Run the Full Pipeline

Run Specific Experiments

How to Run Tests

Documentation Index

Key Documents

Model & Methods

Analysis & Robustness

Guides

Navigating Results

Google Colab Notebooks

Research Questions

Citation

License

Contact

About

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages