Poojitha75/global-welfare-env

🌍 Global Benefit Eligibility & Fraud Detection OpenEnv

🚀 Overview

This project implements an OpenEnv reinforcement learning environment where an AI agent learns to evaluate eligibility for social benefit programs and detect fraudulent claims.

The system simulates real-world decision-making used in government welfare systems globally.

🧠 Key Features

• —✅ Multi-step decision workflow
• —✅ Explainable AI (decision reasoning)
• —✅ Time-based fraud detection
• —✅ Reward shaping with penalties
• —✅ Hybrid decision system (rule-based + LLM-ready)
• —✅ Real-world inspired environment

🔍 Problem Statement

Governments worldwide face challenges in:

• —Incorrect benefit approvals
• —Fraudulent repeated claims
• —Inefficient verification workflows

This environment enables AI agents to learn optimal decision-making policies to address these issues.

⚙️ Environment Design

📊 Observation Space

• —income_level
• —family_size
• —documents_verified
• —previous_claims
• —region_type
• —lastclaimdays_ago

🎯 Action Space

• —approve
• —reject
• —request_verification
• —flagforaudit

🎁 Reward System

• —Correct decision → +1.0
• —Partial decision → +0.5
• —Fraud missed → penalty
• —Step cost penalty → -0.05 per step

🔄 Multi-Step Workflow

1. 1.Request verification (if needed)
2. 2.Final decision (approve/reject/audit)

🧪 Example Output

Example Output

[START] task=welfare env=openenv model=gpt-4.1-mini [STEP] step=1 action=request_verification reward=0.45 done=false error=null [STEP] step=2 action=approve reward=0.95 done=true error=null [END] success=true steps=2 rewards=0.45,0.95 ---

Why This Matters

• —Globally applicable welfare system
• —Fraud detection using behavioral patterns
• —Real-world decision modeling
• —AI-ready environment for RL training

Future Improvements

• —Identity fraud detection
• —Cross-region duplicate claims
• —LLM-based reasoning
• —Dynamic policy learning