theory

Theory Documentation

This folder contains theoretical foundations and explanations for the Bandicoot RMAB system.

Documents

Core Concepts

1. 01-rmab-fundamentals.md - Introduction to Restless Multi-Armed Bandits
2. 02-healthcare-problem.md - The vaccination adherence challenge
3. 03-our-solution.md - Bandicoot's approach and architecture

Advanced Topics

4. 04-whittle-index.md - Understanding Whittle indices and computation
5. 05-clustering-rationale.md - Why clustering over individual learning

Reading Path

For stakeholders/non-technical:

• Start with 02-healthcare-problem.md
• Then 03-our-solution.md
• Skip the mathematical details in 01 and 04

For engineers/researchers:

• Read in order: 01 → 02 → 03 → 04 → 05
• Refer to these when implementing algorithms in /src

For data scientists:

• Focus on 01, 04, and 05
• These explain the statistical foundations

Key Takeaways

The Problem

• 200K+ caregivers, limited health worker bandwidth
• Need to prioritize: who to contact for maximum vaccination impact?
• Random/heuristic allocation is inefficient

The Solution

• RMAB framework models caregivers as bandits with changing states
• Whittle index provides optimal priority scores under budget constraints
• Clustering shares statistical strength across similar caregivers
• Proven approach: SAHELI deployment showed 30% dropout reduction

Why It Works

• Learns from historical data (SMS engagement, vaccinations)
• Adapts to individual behavior patterns
• Balances exploration (uncertain caregivers) vs exploitation (known high-risk)
• Scales to 200K+ caregivers with <$200/month cost

---

Related Documentation:

• Product: /docs/MVP_PRD.md
• Technical Design: /docs/tech-design/
• Archive: /archive/suvita_rmab_chat.md (5,909-line design discussion)