Getting Started with Flatland

Flatland is an environment for developing and comparing multi-agent reinforcement learning algorithms in gridworlds.

This repository contains notebooks to get you started on the right track with the Flatland environment, in order to take part in the AIcrowd Flatland Challenge.

If you want to dive into challenge baselines right away, check out the various approaches below.

Discovering Flatland

Part 1: The Rail Environment

• Create a RailEnv environment and render it
• Check out the default observations
• "Train" a random agent

Part 2: Observations

• Finding suitable observations
• Creating your own observations
• Visualizing observations

Part 3: Level Generation

• Creating random rail networks
• Creating schedules
• Adjusting size and difficulty

Part 4: Malfunctions

• Introducing stochastic malfunctions
• Handling malfunctions

Part 5: Speed Profiles

• Handling agent speed
• Handling partial moves

Challenge Baselines - Coming soon!

The Flatland Challenge can be approached in different ways - for example using methods from operations research, reinforcement learning, or anything else!

The following notebooks show how to approach the problem using each of these methods.

Reinforcement Learning: DDQN

Solve simple rail problems using Double DQN.

Operations Research

Solve simple rail problems using OR methods.