SmartBus-AI

SmartBus-AI allocates bus seats using a MILP optimizer (hard constraints + strong group cohesion) enhanced by ML signals:

• Global models (trained offline from backend’s feedback data)
• Personalized runtime models (tiny per-passenger regressors using their own history)

This repo exposes a FastAPI service (/allocate) and includes an offline trainer to build global models. All data persistence is handled by backend; this service is stateless.

---

Architecture

Optimizer (MILP)

• One seat per passenger.
• No seat reuse when passenger trip intervals overlap.
• Safety zones:
  • Male passengers cannot sit in female-only rows.
  • Non-priority (not disabled & under 60) cannot sit in accessible rows.
  • Wheelchair users must be in accessible rows.
• Objective: strong group cohesion (horizontal > vertical), with optional ML bonuses:
  • Global seat type preference (front/window/aisle).
  • Global penalty for group separation distance.
  • Personalized bonus if the passenger historically rates similar seat locations higher.

ML

• Offline global models (trained with src/train_global_models.py) — saved into models/.
• Runtime personalization per passenger (if history available from backend).

---