Subhajit Paul
>

marl-lab

multi-agent reinforcement learning, training live in your browser — no server, no library, just two Q-tables

episode return (last 200 episodes)

0

episodes

0%

success (last 100)

0

collisions / ep

exploration ε — drag to 0 to watch the learned policy0.25
training speed (steps / frame)150

and are independent Q-learners. Each must reach its dashed goal, but the only way through the wall is the single door — and colliding hurts (−8).

Watch the collisions/ep stat: early on they crash at the door constantly; within a few thousand episodes they learn to stagger their crossing — coordination emerging from nothing but reward.

Under the hood: independent tabular Q-learning (each agent observes both positions, 49×49 states × 5 actions), ε-greedy exploration, α = 0.2, γ = 0.95. The door is the classic bottleneck coordination problem — a toy version of the credit-assignment and coordination questions in my ongoing MARL research. Longer treatment: Albrecht, Christianos & Schäfer, MARL (2024).

Buy Me A Coffee Icon