Reinforcement Learning

Reinforcement Learning is supported by Obit!

Custom Algorithms & Environments

Want to make or implement your own Algorithm and don't see it on the list (e.g PPO, DQN etc)? We extend the Algorithms API for you to implement your own!

We also extend the Environment API to implement custom simulations (e.g pong, CartPole, Pendulum)!

Demo Workloads, Algorithms & Environments

The following RL algorithms are implemented out of the box:

• PPO (for both discrete and continuous action spaces)

• DQN

• Rainbow DQN

• DreamerV2 (partial)

• Custom (user-specified) Algorithm API

The following simulations (environments) are implemented out of the box:

• CartPole

• Mountain Car

• Pendulum

• Pendulum (via the Gymnasium API with the Obit Python Environments API)

• Pong - custom version slightly simplified for speed

• Custom (user-specified) Environment API

The following RL Workloads are implemented out of the box for demo purposes:

• DQN on CartPole

• DQN on MountainCar

• DQN on Pong

• Rainbow DQN on CartPole

• PPO on Pendulum

• PPO on MountainCar (continuous action space)

• PPO on CartPole

• DreamerV2 on CartPole

Autovectorization

Environments can be automatically replicated for a significant speed-up using the Autovectorization API in just a few lines!