Adaptive Traffic Signal Control using PPO Reinforcement Learning

Overview

Intelligent traffic signal controller using Proximal Policy Optimization (PPO) to prioritize emergency vehicles while optimizing traffic flow across six vehicle types.

Key Achievement: 10.6% improvement in emergency vehicle response time

---

Vehicle Priorities

Vehicle Type	Priority	Weight
Emergency	1st	5.0x
Truck	2nd	4.0x
Car	3rd	3.0x
Auto/Taxi	4th	2.0x
Motorcycle	5th	1.0x
Bus	6th	0.5x

---

Performance Metrics

Trinity Network (Primary)

Vehicle Type	Baseline	PPO Agent	Improvement
Emergency	69.00s	61.67s	10.6% ✓
Truck	176.05s	168.57s	4.2% ✓
Car	153.45s	148.30s	3.4% ✓
Motorcycle	148.48s	146.65s	1.2% ✓

Indiranagar Network (Cross-Map Evaluation)

Vehicle Type	Baseline	PPO Agent	Improvement
Emergency	25.73s	23.83s	7.4% ✓
Truck	151.30s	92.39s	39.0% ✓
Bus	90.96s	87.49s	3.8% ✓
Car	82.28s	84.39s	-2.6%

---

Explainable AI (XAI) Results

Reward Decomposition

How the agent allocates optimization effort:

• 45% - Traffic flow optimization
• 30% - Emergency vehicle response time
• 15% - Truck efficiency
• 10% - Car quality of service

SHAP Feature Attribution

Top decision drivers:

• Queue lengths (critical influence)
• Emergency vehicle counts (validates priority learning)
• Lane-specific congestion metrics

Interpretation: Red bars = features promoting "switch signal", Blue bars = features promoting "keep signal"

---

Quick Start

# Install dependencies
pip install -r requirements.txt

# Train the model (30-40 minutes)
python scripts/train_ppo.py --no-gui

# Evaluate performance
python scripts/evaluate_ppo.py --model models/ppo_mg_road/best_model.zip

# Generate XAI visualizations
python analysis/explain_reward.py
python analysis/explain_shap.py

---

System Architecture

Component	Details
Observation	43D (queue lengths, signal phase, vehicle counts)
Action	2 actions (keep/switch signal phase)
Reward	Weighted sum: 45% flow + 30% emergency + 15% truck + 10% car
Network	[256, 256, 128] hidden layers with PPO
Training	150,000 steps, 4 parallel environments

---

Project Structure

scripts/
├── train_ppo.py              # Train the agent
├── evaluate_ppo.py           # Evaluate on test set
├── baseline.py               # Fixed-time baseline
└── visualize_model.py        # GUI visualization

analysis/
├── explain_reward.py         # Reward decomposition
├── explain_shap.py           # SHAP feature attribution
└── reward_analysis.json      # Results

utils/
├── SumoEnv.py               # SUMO environment wrapper
└── ppo_agent.py             # PPO configuration

models/ppo_mg_road/
├── best_model.zip           # Best trained model
└── final_model.zip          # Final checkpoint

logs/
├── ppo_training/            # Training metrics
└── policy_explanation/      # SHAP plots & analysis

---

Key Findings

✅ 10.6% emergency response improvement (Trinity)
✅ 39% truck performance improvement (Indiranagar)
✅ 7.4% emergency improvement (cross-network)
✅ Learned to prioritize emergency vehicles correctly
✅ Generalizes to different traffic networks

---

Status: Production Ready | Version: 1.0 | Updated: February 2026