Flatland Challenge: Multi-Agent Reinforcement Learning on Trains

We are running a NeurIPS challenge where the goal is to schedule trains using RL.

It tackles a real-world problem: railway networks are growing fast, but the classical decision-making methods used today don’t scale well. This is becoming problematic! Can RL save the day?

Our goal is to foster research in RL around this problem, and to establish a benchmark showing the progress of RL against other (currently better!) methods.

We are hoping for an “AlphaGo moment”, where reinforcement learning will take over. Planning train schedules actually has many similarities with the game of Go!

We provide strong baselines and "getting started" guides to help you hit the ground running, even if you're just starting with RL. For example you can run this Colab notebook to train a DQN policy that you can then submit straight to the leaderboard: https://flatland.aicrowd.com/getting-started/rl.html