Agentic Inventory Optimizer

Multi-Agent Reinforcement Learning for Multi-Warehouse Inventory Management

📄 Document & Video

📂 Full Project Report & Demo Video

---

📌 Overview

This project implements an Agentic Workflow System for optimizing multi-warehouse inventory operations using Reinforcement Learning (RL).
It integrates:

• Value-Based Learning → Deep Q-Network (DQN)
• Policy Gradient Methods → Proximal Policy Optimization (PPO)
• Exploration Strategies → Upper Confidence Bound (UCB) policy selection
• Custom Agentic Tools → Cost simulation, decision explanation, and warehouse Q&A

The goal is to minimize total inventory costs (holding + stockout + ordering) while maintaining high service levels under uncertain demand.

---

🚀 Features

• Two RL approaches: DQN & PPO
• Multi-agent orchestration with policy selection
• Fallback mechanism for safety
• Custom tools for:
  • Cost simulation
  • Decision explanation
  • Warehouse Q&A
• Streamlit dashboard for visualization
• Baseline comparisons with heuristic policies
• Exportable reports with learning curves & performance breakdowns

---

📂 Repository Structure

agentic-inventory-optimizer/
├── agents/                  # RL agents, forecaster, orchestrator, policy selector
├── custom_tools/             # Cost simulation, dashboard export, decision explainer, QA
├── env/                      # Inventory environment & wrappers
├── demo/                     # Streamlit UI
├── rl/                       # Training & evaluation scripts
├── results/                  # Models, evaluation JSONs, and visualizations
├── tests/                    # Unit tests for agents & tools
└── README.md                 # This file

---

⚙️ Installation

1. Clone the repo

git clone https://github.com/anumohan10/agentic-inventory-optimizer.git
cd agentic-inventory-optimizer

2. Create virtual environment & install dependencies

python -m venv venv
source venv/bin/activate   # Linux/Mac
venv\Scripts\activate      # Windows
pip install -r requirements.txt

---

▶️ Usage

1. Train an RL Agent

DQN

python -m rl.train_rl_agent --algo dqn --episodes 10000     --target_service 0.92 --below_target_mult 12.0 --seed 42

PPO

python -m rl.train_rl_agent --algo ppo --episodes 8000     --target_service 0.92 --below_target_mult 8.0 --seed 0

2. Evaluate a Model

python -m rl.evaluate_agent --algo ppo --episodes 100     --model_path results/models/ppo_best_98_service.zip

3. Run the Dashboard

streamlit run demo/app.py

---

📊 Results

Policy	Total Cost	Service Level	Notes
PPO	$6,504	98.85%	🥇 Best
DQN	$6,897	98.33%	🥈 Excellent
Heuristic	~$7,500–8,500	85–92%	📊 Baseline

---

🏗 Architecture Diagram

---

📌 Key Achievements

• DQN improved service from 69.7% → 98.33%
• Cost reduction of 39% for DQN after tuning
• PPO achieved optimal cost-service trade-off
• Agentic orchestration with policy switching and fallback safety

---

📈 Future Improvements

• Multi-agent RL (one per warehouse)
• Continuous action spaces
• Integration with real demand forecasting models
• Transfer learning between warehouses
• Testing with real-world supply chain datasets

---

📜 License

MIT License © 2025 [Anusree Mohanan]