pufferlib

Vectorized environments accelerate reinforcement learning training by parallelizing environment execution. This toolkit provides a high-performance framework supporting single-agent and multi-agent workflows across Gymnasium, PettingZoo, and Ocean ecosystems for rapid experimentation.

Multi-GPU reinforcement learning training with torchrun is supported by this framework. It requires PyTorch and the PufferLib package to scale high-throughput PuffeRL-based training across multiple GPUs for single-agent and multi-agent tasks.

Multi-agent reinforcement learning supports flexible policy architectures including MLP, CNN, LSTM, and multi-input networks. The framework provides built-in initialization patterns and debugging utilities for training agents across PettingZoo and Ocean environments.

Bootstrapping a reinforcement learning experiment with Gymnasium uses the provided train_template.py script. This template initializes a PuffeRL experiment against an Ocean or Gymnasium environment to rapidly prototype and compare agents.

PufferLib works seamlessly with PettingZoo environments for multi-agent training. It integrates with the PettingZoo ecosystem to prototype, scale, and compare agents using vectorized environments and high-throughput training.

Scaling single-agent reinforcement learning for benchmarking is best handled through high-throughput vectorized training. This framework streamlines environment management and policy optimization to rapidly experiment and compare agents across Gymnasium and Ocean.