INDEPENDENT MODULE / 2.0

Build the experiment.
Then search the settings.

49 inputs. Eight intersection designs. A visible protocol for every result, with fresh random seeds to test the settings the search selects.

Settings compete within the same traffic, vehicle and cost scenario. Changing demand or assuming cheaper fuel does not count as improving the intersection. The objective is fuel + full-trip vehicle time; crash risk and construction cost are not optimized.

01 / DEFINE THE EXPERIMENT

Choose a starting point, then edit every range.

Each sweep accepts comma-separated values or a numeric range start:stop:step. Shares are fractions: 0.2 means 20%. Fixed values also define the baseline used in validation. Sampled search is reproducible coverage of a grid; it is not exhaustive.

Crash and total-cost scope

This module does not infer how changing a signal, gap or geometry changes crashes. The comparison lab values user-supplied whole-system crash rates and KABCO severity mixes separately, using FHWA costs. The original two-layout lab retains its stochastic crash process. Neither provides a calibrated safety response for this search.

Clearance times, gap acceptance, speed and vehicle properties are sensitivity assumptions, not values to minimize. Local red-turn permissions and safe clearance times require site-specific review. Queue storage is a diagnostic sampled every 10 seconds; physical spillback and pedestrian conflicts are not simulated.

02 / COVERAGE & VALIDATION

Set the compute budget before seeing the results.

Search modeSampled setting draws / designSampled assumption scenariosScreening replicationsFresh validation replicationsScreen seedValidation seedGrid sampling seedAssumed dollar year

Exact mode tests every distinct applicable combination, with limits of 10,000 effective candidates and 100,000 screening simulations per batch; it never silently samples. Periodic offset aliases and inactive settings are deduplicated. In sampled mode, removing duplicates can leave fewer candidates than draws. Valuation cases reuse traffic outcomes. No sweep starts automatically.

03 / FROZEN CHOICES & FRESH TESTS

See what held up.

Ready to define an experiment.

Run a search to compare each design’s frozen screening choice with its unchanged baseline. No result has been selected in this build.

Assumption scenarioValuation

Design / frozen choice	Baseline / 1k trips	Selected / 1k trips	Paired Δ ± 95% seed interval	Annual Δ	Validation / settings	Replay

Fresh means do not choose a new winner. Comparisons across designs are descriptive; the frozen choice was selected separately within each design. Intervals measure simulation seed variation only, not input, structural or search uncertainty. Annualization scales all affected traffic to the assumed intended-left volume; it is not a measured statewide saving.

Search audit and rejected candidates

Scenario	Design	Candidate	Screen cost / 1k trips	Eligibility	Inputs

On-screen audit shows the first 100 records for the selected case; JSON and CSV contain every record.

Build the experiment.Then search the settings.

Choose a starting point, then edit every range.

Set the compute budget before seeing the results.

See what held up.

Build the experiment.
Then search the settings.