NAGA-SIDDHARDHA/theft-rl

RL-Based Theft Vehicle Monitoring System

Problem

Detecting stolen vehicles efficiently under limited surveillance resources.

Approach

We model this as a Reinforcement Learning problem where an agent decides:

• Ignore
• Track
• Alert

State

• Suspicion score (0–10)
• Available tracking resources
• Tracking status

Actions

• 0: Ignore
• 1: Track
• 2: Alert

Reward Design

• +20 → Correct alert
• -10 → False alert
• -15 → Missed stolen vehicle
• +10 → Useful tracking
• Resource penalties included

Key Idea

Tracking improves the quality of information (reduces uncertainty), allowing better future decisions.

Result

The agent learns to:

• Focus on high-risk vehicles
• Avoid unnecessary alerts
• Efficiently use limited resources

How to Run

python train.py
## Results

The agent shows improvement over time:

- Initial episodes: low/negative rewards
- Later episodes: higher rewards

This indicates the agent learns better decision-making policies.
# sample results
First 20 epoches avg reward: -8.4
Last 20 epoches avg reward: 48.9