divyasridoredla/traffic-incident-resilient

Incident-Resilient Traffic Signal Control

Problem

Standard RL traffic agents collapse under real-world disruptions. This project builds a custom Gymnasium environment that injects disruptions mid-episode, training a robust PPO agent.

Disruption Types

• Lane closure — obs[0] += 0.5
• Demand shock — all obs += 0.3
• Sensor dropout — obs[2:] = 0

Results

Agent	Normal	Under Disruption
Baseline PPO	-17.85	-73.22
Robust PPO	-22.71	-60.28

17.8% improvement under disruption conditions.

Module Mapping (Hackathon)

• Module 1: Custom env required — no standard env models mid-episode disruptions
• Module 2: Baseline PPO trained on TrafficEnv (no wrapper) as benchmark
• Module 3: Registered as IncidentTrafficEnv-v0 via gym.register()
• Module 4: TrafficEnv(gymnasium.Env) + DisruptionWrapper(gymnasium.Wrapper)

How to run

pip install gymnasium stable-baselines3 fastapi uvicorn
python -m uvicorn main:app --port 8000
python inference.py