cbi-plus 2.12.0

CBI+ — Congestion Bottleneck Identification / FD / QVDF per-corridor calibration pipeline (queue objects T0/T2/T3, discharge rate mu, capacity drop, QVDF elasticities) for PeMS and INRIX TMC data

CBI+ — Congestion Bottleneck Identification / FD / QVDF calibration pipeline

In plain words: this tool diagnoses WHY a freeway is congested — which bottleneck causes it, when it starts, how long it lasts — instead of just plotting that it is. (AMS = Analysis, Modeling & Simulation of traffic.)

Agentic AI for Translational AMS Modeling — first working instance. Scheme-driven · engine-agnostic · quality-gated.

New here? Two five-minute reads before anything else: docs/GLOSSARY.md (every symbol, plain English) and docs/SETUP_FOR_BEGINNERS.md (where to type the commands, how to open a notebook).

The paradigm in one frame — left: plot-driven traffic engineering (PHF, V/C, hourly averages — reports numbers, cannot explain or decide). Right: mechanism-based diagnosis (5-minute dynamics, cumulative-curve queue physics T0/T2/T3, μ/C capacity drop, QVDF reconstruction, benchmark gates) feeding real decisions. Every element on the right is runnable in this repository.

What this repo has already proven (all regenerate with one command; the gates are internal reproduction checks, not external peer review):

25/25	5	12	7	2	5
reproduction gates PASS	published-work reproductions (Tables 5/6/7 exact)	engines, one contract	corridors diagnosed	arenas where engines compete	executed teaching notebooks

Front doors: 🌐 website front page · 📇 onboard a corridor / add a dataset (incl. parquet) · 🔤 add vocabulary (tokens/issue types) · 🗣️ CBI tokens — the planner-facing vocabulary · 📓 teaching notebooks (start with 01 — no data needed) · 🔬 benchmark reproductions · 📦 pip package guide · 📖 the Introduction

A general-purpose, per-corridor calibration toolkit for the four-layer traffic-engineering pipeline:

Congestion Bottleneck Identification → Fundamental Diagram → discharge-rate μ validation → QVDF (Queue Volume-Delay Function) forward model

This tool is not about learning how to run another traffic software package. It is about learning how to think like a modern traffic engineer: observe, diagnose, explain, validate, and decide.

# Type these in a terminal: Command Prompt on Windows, Terminal on Mac.
# Never used one? -> docs/SETUP_FOR_BEGINNERS.md
pip install cbi-plus            # (or `pip install -e .` from this repo folder)

from cbi_pipeline import api
df = api.simulate_corridor(days=5)   # no data files needed — planted AM bottleneck
out = api.diagnose(df)               # QC -> episodes -> FD -> QVDF -> CBI ranking
out["ranking"].head()                # who deserves the money, and why

Start here: docs/INTRODUCTION.md — who we train, what the old teaching misses, the LWR–Newell–Daganzo foundation, and the mission.

Why this platform exists: docs/MISSION.md (the medical-school analogy and the seven operating principles) · theory foundations: docs/teaching/THEORY_FOUNDATIONS.md (why V/C is not congestion; LWR-Newell-Daganzo). New here? Read docs/GLOSSARY.md first (5 minutes - every symbol, and the paper-to-pipeline name map).

Given a corridor of detectors (speed + optional volume, 5-minute cadence), CBI+ answers, per detector × day × period (AM / MD / PM):

Question	Output
When did congestion start / peak / clear?	queue object T0, T2, T3 + duration P
How bad?	lowest speed v_t2, regime ∈ {uncongested, mild, recurring, severe, event}
What does the bottleneck actually serve?	discharge rate μ = median flow while the queue drains
How much capacity was lost?	μ/C — the capacity drop (typically 0.85–0.95)
Can two numbers reproduce the day?	QVDF elasticities (Q_n, Q_s, Q_cd, Q_cp) + closed-form v(t)
Is any of this trustworthy?	quality gates + per-episode audit panels (PNG)

Handles both PeMS (speed + measured volume → real S3 FD fit) and INRIX TMC (speed-only → volume synthesized via the CBI inverse-S3 prior).

Install / run

pip install -e .        # installs cbi_pipeline + all dependencies (incl. scipy)

# Your first 30 minutes (every step runs on IN-REPO data):
python -m cbi_pipeline.repro_gates                 # 0. is my clone healthy? (~5 s)
# 1. read docs/GLOSSARY.md (5 min) then docs/STAGE_CHAIN.md (the one-page map)
# 2. the hello world (reproduces the QVDF paper exactly, ~2 min, in-repo data):
#      cd benchmarks/qvdf_paper_i10 && python reproduce_qvdf_paper.py   # I-10 (Case Study 1)
#    or the I-405 flagship (Case Study 2):
#      cd benchmarks/qvdf_paper_casestudy2 && python reproduce_casestudy2.py
# 3. full pipeline on in-repo I-10:  python teaching_cases/case_02_ca_pems_i10.py

# CA PeMS corridor (measured volume)
python -m cbi_pipeline.corridor_workflow --corridor 10-E --source pems --pems-path benchmarks/I-10/link_performance.json

# AZ INRIX TMC corridor (speed-only, CBI inverse-S3 synthesis)
python -m cbi_pipeline.corridor_workflow \
    --corridor I-17 --source inrix \
    --inrix-folder additional/benchmark_datasets/datasets/I-17 \
    --s3-prior az_tmc_i17 --rederive-kc-and-m

Outputs per corridor: stage1_qc/ … stage5b_corridor/, FIXED-layout figures/, per-episode verification panels/, and quality_gates.json (PASS/FAIL).

TrafficFlowBench (IEEE competition) bridge

tfb_adapter.py runs the full pipeline directly on the TrafficFlowBench PeMS-LA release (parquet detector states + GMNS network):

# REQUIRES the external TrafficFlowBench release (not in this repo):
#   set TFB_DATA_ROOT=<path to 02_data_PeMS_LA> - nothing else here needs it
python tfb_adapter.py I-210E 2026-06-01 2026-06-28
python tfb_teaching_extract.py     # emit the CBI-Lab teaching payload (data.json)

The adapter derives effective lanes from the data (never trust map lane tags), respects the release's is_observed imputation mask, and feeds the standard pipeline. The extracted payload drives the CBI Lab interactive teaching page (gui4gmns → TrafficFlowBench → CBI Lab): raw space-time field → identified queue objects → physical queue under the microscope → corridor law.

Start with docs/TFB_CBI_GUIDE.md — how to run, what to read in which order, healthy magnitudes, per-period interpretation, pitfalls.

The five stages

Stage	Module	Produces	Outlier mitigation
1 QC	stage1_qc	qc_pass, cleaned speed	Hampel MAD, jump, spatial-neighbor, per-day multi-bottleneck wave-direction check
2 Episodes	stage2_episodes	queue objects per (sensor, day, period) + MD→PM boundary merge	per-(sensor, period) z-score event flag; NaN-gap + persistence hardened scan
2b Screen	stage2b_measured_diagnostics	physical-violation flags on measured D/C, μ, V_t2	Huber residuals
3 FD	stage3_fd_robust	per-sensor S3 fit (capacity, v_c) + bootstrap CI	Huber loss, regime-separated
4 μ	stage4_mu_validation + stage4_verification	μ per episode/link + step-by-step audit CSV/panels	discharge-window median, group shrinkage
5 QVDF	stage5_qvdf + stage5_verification + stage5b_corridor_aggregate	elasticities per (sensor, period) + exact round-trip audit + corridor law with bootstrap CI & prior shrinkage	IQR, feasibility ranges (C++ verbatim)

Closeout: benchmarks, gates, and the CBI ranking

CBI+ is accepted only against evidence — see docs/CLOSEOUT.md:

• Stage 6 (CBI deliverable): every run emits stage6_cbi/benchmark_bottleneck_ranking.csv — sensor-period bottleneck scores (frequency × duration × severity) with active / passive / spillback classes and explicit aggregation-level labels.
• Hard gates: python -m cbi_pipeline.benchmark_gates <run_dir> writes pass_fail_summary.csv + benchmark_comparison_report.csv (physics bands, episode sufficiency, QVDF round-trip ≤ 10 mph MAE, ranking stability vs a reference run). Thresholds: benchmarks/benchmark_validation_tolerance_template.csv.
• Benchmark cases in-repo: benchmarks/I-10 and benchmarks/I-405 (public Caltrans PeMS, March 2018, with their own sensor tables + MANIFEST.json) reproduce with one command each; I-395 NVTA runs from the private data path.
• Dataset discovery: python -m cbi_pipeline.discover <root> inventories every CBI-compatible dataset under a directory tree with readiness verdicts.

Provenance & audits

Ten issues were found and fixed during the 2026-07 TrafficFlowBench integration (period-relative indexing, D/C units, imputed-data ingestion, single-bottleneck direction gate, …) — every one documented with evidence and root cause in docs/FIXES_CBI_PLUS_2026-07-07.md. Read it before comparing against pre-fix outputs. The design lesson running through all ten: physics quantities get exactly one implementation, indices carry their frame with them, and priors (maps, imputation) are never treated as measurements.

Teaching cases

Four self-contained scripts under teaching_cases/ (AZ I-17 INRIX week, CA I-10E and I-405S PeMS months, and a cross-corridor comparison). See teaching_cases/README.md.

Lineage

Successor to clean_handoff_v1/v2 (Abbasi & Zhou, ASU) and the original CBI tool; QVDF forms mirror the DTALite C++ scan_congestion_duration / calculate_travel_time_based_on_QVDF flow verbatim, including feasibility ranges. Companion visual layer: gui4gmns.

Community positioning

We are all friends in the broader traffic-engineering, AI, and data-science community. Different groups bring different perspectives: theory, operations, dashboards, machine learning, simulation, agency practice, and open-source engineering. Our goal is to learn from those perspectives and make the tool more reliable, explainable, and useful.

We are not competing against individual scholars, labs, or tools. We are trying to improve a shared workflow: from static plotting to mechanism-based, data-driven, benchmark-validated traffic engineering.

How we practice this: simulated-participant audits run against every release (docs/AI_PARTICIPANT_PROTOCOL.md); every finding lands in the public issue register with run/issue logs in docs/reviews/.

License, citation, scope

MIT — see LICENSE. Cite via CITATION.cff (the QVDF methods paper: Zhou et al. 2022, Multimodal Transportation 1:100017). Maintained by the ASU Trans+AI Lab. Current scope: US freeway corridors (PeMS / INRIX / IEEE TrafficFlowBench data); units are US customary (mph, veh/mi) with conversions stated at every loader — apply judgment before generalizing elsewhere. Issue history and enhancement backlog: docs/ISSUE_REGISTER.md.