Multi-Agent Reinforcement Learning — Delivery Routing

Multi-agent reinforcement learning project using tabular Q-learning for collaborative package delivery routing.

Overview

This project implements a multi-agent RL system on a 5×5 grid where four agents learn to coordinate package deliveries between locations A and B. Each agent keeps its own Q-table over the discretized joint state and learns independently while sharing a common environment clock.

Key features:

• Tabular Q-Learning — one Q-table per agent over the joint state (agent positions and delivery phases)
• Epsilon-Greedy Exploration — decaying exploration rate for convergence
• Hybrid Reward Shaping — movement penalty, stationary penalty, collision penalty, inverse-distance guidance, and delivery bonuses
• Collision Detection — penalties when agents in opposite delivery phases meet outside the goal cells
• Centralized Clock with Simultaneous Learning — shared environment step counter

Files

• Multiagent_RL_Delivery.ipynb — full project notebook with training, evaluation, and analysis
• DeepQ_Skeleton.ipynb — separate exploratory DQN skeleton, not used by the main notebook

Getting Started

pip install numpy matplotlib
jupyter notebook Multiagent_RL_Delivery.ipynb

Results

Focal-agent evaluation covers 2 delivery maps × 16 start configurations (all 2⁴ phase assignments). All four agents reach a 100.00% success rate, where success requires completing the round trip with zero collisions.