Hackable PyTorch RL Library with Distributional Algorithms (D4PG, DSAC, DPPO) (github.com)

A new hackable PyTorch library focused on reinforcement learning (RL) has been released, featuring efficient implementations of several advanced algorithms designed to run entirely on GPU. The library includes key algorithms such as Distributed Distributional DDPG (D4PG), Distributional PPO (DPPO), and Soft Actor Critic (SAC), making it a versatile tool for researchers and developers in the AI/ML community. Users can easily install the package and choose additional functionalities for enhanced performance, including integration with Gymnasium and TensorBoard. This library is significant because it streamlines the development and experimentation process for cutting-edge RL methods, allowing for rapid prototyping and training of algorithms on different compute backends, including CUDA and Apple Silicon. The built-in device resolution helps users optimize performance without needing deep technical knowledge. Additionally, the project includes pre-tuned hyperparameters and examples, such as the demonstration of policy convergence for the BipedalWalker-v3 environment, showcasing its practical application in training RL models effectively. Researchers are encouraged to cite the work and contribute to its ongoing development, which highlights the collaborative nature of the field.