Inventory Quantum

Advanced Inventory Optimization with Deep Learning, Quantum Computing, and Reinforcement Learning

---

🎯 Overview

A research-grade inventory optimization system demonstrating:

Component	Technology	Innovation
Forecasting	TFT, LightGBM, ARIMA	Interpretable attention + Conformal intervals
Optimization	QUBO + QAOA	Classical quantum simulation
RL	Conservative Q-Learning	Offline policy learning
Monitoring	PSI, KS-test	Distribution-free drift detection

---

✨ Key Features

📈 Deep Forecasting

• Temporal Fusion Transformer (TFT) - Google's SOTA model (2021)
• Variable Selection Networks - learned feature importance
• Conformal Prediction - distribution-free coverage guarantees

⚛️ Quantum-Inspired Optimization

• QUBO Formulation - binary encoding for order quantities
• QAOA Simulation - variational quantum eigensolver
• Hybrid Solver - automatic fallback for large problems

🤖 Offline Reinforcement Learning

• Conservative Q-Learning (CQL) - prevents OOD overestimation
• Off-Policy Evaluation - IS, WIS, FQE methods
• No online interaction required

🔍 Drift Detection

• Population Stability Index (PSI)
• Kolmogorov-Smirnov test
• Jensen-Shannon divergence
• No ground truth labels needed

---

🚀 Quick Start

# Clone
git clone https://github.com/username/inventory-quantum
cd inventory-quantum

# Docker (recommended)
docker-compose up --build

# OR local installation
pip install -r requirements.txt
python -m uvicorn app.main:app --reload

API docs at: http://localhost:8000/docs

Dashboard at: http://localhost:8000/static/dashboard.html

---

📊 API Endpoints

Endpoint	Method	Description
/upload/	POST	Upload CSV data
/forecast/{sku}	POST	Generate 30-day forecast
/forecast/compare	POST	Compare models
/decision/optimize	POST	Get reorder recommendation
/decision/qubo/ablation	GET	QUBO vs classical comparison
/monitoring/drift	GET	Check distributional shift

---

🔬 Research Highlights

Theoretical Contributions

• Conformal Coverage: P(Y ∈ [L, U]) ≥ 1 - α (finite-sample)
• QUBO Encoding: O(n) binary variables for quantity levels
• CQL Regularization: Prevents value overestimation on OOD actions

Experimental Results

Model	RMSE	Coverage 95%
Naïve	8.45	78.3%
ARIMA	6.32	82.1%
LightGBM	5.18	88.5%
TFT + Conformal	4.92	94.8%

Statistical Rigor

• Friedman test with Nemenyi post-hoc
• Bootstrap confidence intervals
• Shapley values for feature attribution

---

📁 Project Structure

inventory-quantum/
├── app/
│   ├── main.py              # FastAPI app
│   ├── forecasting/         # TFT, ARIMA, LightGBM, Conformal
│   ├── decision/            # QUBO, QAOA, policies
│   ├── rl/                  # CQL, policy evaluation
│   ├── monitoring/          # Drift detection
│   └── evaluation/          # Statistical benchmarking
├── docs/
│   ├── technical_report.tex # NeurIPS format
│   └── RESEARCH_SUMMARY.md
├── tests/
├── Dockerfile
└── docker-compose.yml

---

📚 References

1. Lim et al. (2021) - Temporal Fusion Transformers
2. Romano et al. (2019) - Conformalized Quantile Regression
3. Farhi et al. (2014) - QAOA
4. Kumar et al. (2020) - Conservative Q-Learning

---

📄 License

MIT License - See LICENSE for details.

---

Built for MIT / Stanford / CMU application portfolio