Optimizing Traffic-Signal Timing on Big Bend Boulevard:
Data, Simulation and Validation
Summary. 4 County signals on Big Bend Blvd · weekday 5–6 PM · real County timing + counts · SUMO simulation validated ({ok}/{len(inter)} GEH < 5) · County plan {cur:.0f} veh-h/h → model plan {fo:.0f} ({100*(1-fo/cur):.0f}% less) → simulator-optimized {sl:.0f} ({100*(1-sl/cur):.0f}% less). Preliminary: optimizer is our stand-in for the group's code.
1. Corridor and scope
4 signals, north → south · main road Big Bend (NB/SB), cross streets (EB/WB) · design hour weekday 5–6 PM · {tot:,.0f} veh/h total approach volume (summed over the 4 intersections).
Table 1. Links between consecutive signals (length from County coordinates; posted speeds from MoDOT 2025).
2. Data
2.1 Sources
{tbl(["source", "what we use", "status"], [ ["St. Louis County timing sheets (4 signals)", "PM pattern 37: cycle, splits, offsets, yellow/all-red, phase sequence", "measured"], ["St. Louis County turning counts (4 signals)", "5–6 PM volumes by approach + L/T/R · Forsyth 2017, FPP 2021, Wydown 2025, Clayton 2026", "measured · verified vs sheet totals"], ["MoDOT 2025", "lanes, speed limits", "measured"], ["County signal inventory", "signal IDs, ownership, coordinates", "measured"], ["OpenStreetMap", "approach lanes at the stop line", "measured"]])}Every number is tagged measured / derived / assumed (appendix).
2.2 PM peak-hour volumes
{tbl(["intersection", "approach", "veh/h", "L", "T", "R", "lanes", "source"], vol)}Table 2. Measured approach volumes and turning shares, 5:00–6:00 PM. Lanes are approach lanes at the stop line.
2.3 Current signal timing (County, PM pattern)
{tbl(["intersection", "sheet", "pattern", "cycle (s)", "offset (s)", "phase splits (s)", "Big Bend block", "cross block", "lost time L (s)"], tim)}Table 3. County timing, weekday PM. Splits include yellow and all-red. Lost time is the total clearance along the phase sequence.
3. Method
3.1 Pipeline
{tbl(["step", "what", "input", "output"], pipeline)}Table 4. One command (./run_all.sh) reruns all steps. Simulation + Code: live mini-simulation and every program →
3.2 Timing model
Variables: cycle C, Big Bend green gi, offsets · objective: Webster + platoon delay · constraints: 60 ≤ C ≤ 150 s, min greens, x ≤ 0.9 · solver: multistart SLSQP.
3.3 Simulation
{tbl(["setting", "value"], [ ["software", "SUMO 1.28"], ["network", "4 signals at real spacing · real approach lanes · left-turn bays · posted speeds"], ["demand", "measured volumes + turn shares · mid-block balancing flows (Table 5)"], ["signals", "County phase sequence (incl. left arrows), scaled to each plan · right turn on red"], ["run", "15 min warm-up + 60 min measured · 10 seeds per plan"], ["metric", "delay = time lost on corridor links + time waiting to enter (veh-h per hour)"]])} {tbl(["link", "direction", "propagated veh/h", "counted veh/h", "balancing veh/h"], bal)}Table 5. Mid-block balancing flows: counted volume into a signal minus the volume propagated from the previous signal.
4. Validation
Simulated vs counted volumes, GEH < 5 = pass · interior approaches: {ok}/{len(inter)} pass.
{tbl(["intersection", "approach", "kind", "counted veh/h", "simulated veh/h", "GEH"], val)}Table 6. Simulated vs counted volumes under the County's current timing (mean of 10 seeds).
5. Results
{tbl(["plan", "cycle (s)", "SUMO delay, veh-h/h (mean ± sd)", "vs County", "stops / veh", "throughput veh/h", "formula delay", "source"], res)}Table 7. All plans, 10 seeds each. formula-optimum-standin = our stand-in for the group's optimizer.
6. Study 1: simulator-in-the-loop vs formula model
{tbl(["test", "result"], [ ["Does the formula rank plans like SUMO? ({s1['prediction']['n_plans']} random plans)", "weakly · Spearman ρ = {s1['prediction']['spearman']:.2f}"], ["Formula optimizer", "{fo:.0f} veh-h/h · {s1['formula_optimizer']['seconds']} s"], ["Simulator-in-the-loop (Nelder–Mead, {s1['sumo_loop']['evaluations']} evals)", "{sl:.0f} veh-h/h · ~{s1['sumo_loop']['seconds']//60} min"]])}7. Interfaces for the optimizer
In: optimizer_input.json (q, s per approach; τ per link; parameters) · Out: plan file:
{{ "name": "...", "C": 124.7,
"lights": {{ "051": {{"g_main": 37.0, "offset_s": 0.0}},
"040": {{"g_main": 49.8, "offset_s": 52.7}}, ... }} }}
g_main = Big Bend green (s); cross-street green = C - L_i - g_main
offset_s = start of the Big Bend block relative to Forest Park Pkwy (s)
8. Modelling choices for the group to confirm
- Lost time + capacity computed along the real signal sequence (L ≈ 21–27 s) · needs Michael + Noah's OK · switchable back.
- Real signals use left-turn arrows · simulation includes them · revisit "future work".
- Simulated network generated from measured geometry (full OpenStreetMap version jammed).
- Assumed: 1,800 veh/h/lane · p = 0.8 · 90 m left bays · right turn on red.
9. Status
{tbl(["item", "status"], [[a, b] for a, b in status], "status")}10. Reproducing the results
cd pipeline ./run_all.sh # data, network, plans, 10-seed runs, validation, tables .venv/bin/python src/study1.py --n-random 30 --budget 120 # Study 1
Inputs: data/, config/ · outputs: build/ (regenerated) · Simulation + Code: all code, settings and data, plus a zip →