ExploreRL

Train, visualize, and experiment with Reinforcement Learning algorithms entirely on-device with ExploreRL.

Built with MLX, ExploreRL provides an interactive sandbox to test various environments and algorithms with different hyperparameters to observe how RL agents learn through trial and error.

LEARN BY DOING

Train agents on classic adapted environments and observe learning in real time. Watch values grow, track reward curves, and develop intuition for how different algorithms and hyperparameters affect performance. Read and learn about reinforcement learning concepts in the documentation.

ALGORITHMS

Four common implementations:

- Q-Learning

- SARSA

- Deep Q-Network (DQN) for discrete action spaces

- Soft Actor-Critic (SAC) for continuous control

INCLUDED ENVIRONMENTS

Adapted from classic standards:

Toy Text: Frozen Lake, Blackjack, Taxi, Cliff Walking

Classic Control: CartPole, Mountain Car, Mountain Car Continuous, Acrobot, Pendulum

Box2D: Lunar Lander, Lunar Lander Continuous, Car Racing, Car Racing Discrete

FEATURES

- Environment-specific settings

- Real-time training metrics and reward visualization

- Configurable hyperparameters: learning rate, discount factor, epsilon decay, batch size

- Agent library with save, load, duplicate, and export functionality

- Evaluation mode to test trained agents without further training

- Educational documentation covering RL concepts, neural networks, and algorithms

- Custom seed for reproducibility

Note: Training performance will vary by device. Devices with newer hardware and more memory will have a better experience.