Intersection Lab 0.4.0

Changes

• Training counts every delayed person. The reward now includes vehicles held at the edge of the map and pedestrians waiting to cross, not just vehicles on the road, so a controller can no longer starve an approach or a crosswalk for free.

• Honest model selection. Checkpoints are chosen on 16 validation scenarios (design demand, lighter and heavier traffic, randomized demand, bad weather, failed loop detectors) and then scored on 24 separate test scenarios that no choice ever uses. A checkpoint is accepted only if its pedestrian waits stay within 5 s of the predictive controller's.

• Deep RL protects pedestrians. Its safety shield now gives the next walk to any pedestrian who has waited 40 s. On the held-out test set (mean of three training seeds), pedestrian waits fall by 4 to 26 s against 0.3.0, most on the T-junction (45.6 → 19.7 s). Person delay is about the same on three layouts and 2.1 s lower on major/minor. On the four-way layouts the rule also lowers vehicle delay.

• Both AI controllers retrained with one hidden layer of 128 units for 800 episodes (train all 800 900 dqn and train all 800 900 hybrid reproduce them). The hybrid now beats predictive control on every layout's test set, by 0.4 to 1.3 s. Most of that comes from a better fixed tuning of the planner; its network adds to that on the arterial and town layouts.

• Benchmark person delay (bench all 20 6):

  	Arterial	Major/minor	Town	T-junction
  Hybrid, 0.3.0	17.9 s	21.5 s	22.4 s	13.4 s
  Hybrid, 0.4.0	16.8 s	21.5 s	21.6 s	13.7 s
  Deep RL, 0.3.0	19.4 s	24.4 s	30.0 s	11.7 s
  Deep RL, 0.4.0	21.3 s	26.0 s	29.0 s	12.3 s

  The benchmark scores one network per layout at design demand. Networks trained from different seeds differ by 1 to 3 s, so the README's held-out comparison across three seeds is the fairer measure of Deep RL's change.

• Accessibility. Dropdowns can be opened and used from the keyboard, Escape closes the About and benchmark windows, and Space pauses only when no control has focus. Dropdowns are named for screen readers in the desktop app. Small text is larger (11 px), and all panel and chart text meets 4.5:1 contrast. ACCESSIBILITY.md describes what works and the known barriers.

• Training options: RESUME=1 must still beat the best fixed tuning before saving, and SEED and HIDDEN set the seed and hidden layers.

• Project: a code of conduct, contributing guide, security policy (with private vulnerability reporting), issue forms and a pull request template. Windows signing moved to the renamed Azure Artifact Signing action.

Downloads

Signed and notarized installers for macOS (DMG), signed installers for Windows x64 and arm64 (MSI, ZIP), and Linux packages (deb, rpm, tar.gz) for x86_64 and aarch64. Checksums are in SHA256SUMS.txt. The browser version is on the project's GitHub Pages site.

Full Changelog: v0.3.0...v0.4.0