RLlib Baselines

Note

Looking for simpler RL baselines? Checkout the DQN examples implemented from scratch using PyTorch.

The following baselines provide a starting point to develop advanced reinforcement learning solutions. They use the RLlib framework, which makes it easy to scale up training to larger machines or even to clusters of machines.

🔗 RLlib Baseline Repository

Follow the getting started guide to setup and start training using the RLlib baselines.

The RLlib documentation provides ample information about the standard methods such as Ape-X and PPO. The documentation below gives more details about the methods, custom observations and other approaches that have been developed specifically for Flatland.

RL methods

• Centralized Critic PPO

• Imitation Learning: MARWIL, Ape-X DQfD

Custom observations

• Density observations

• Combining observations

Other approaches

• Frame skipping

• Action masking