nvlabs/gbrl

GBRL is a machine learning library from NVIDIA that lets you use tree-based models for reinforcement learning, instead of the neural networks that power most systems today. It is designed for scenarios where an AI agent needs to learn by trial and error, like navigating a robot or playing a game. Most reinforcement learning systems rely on neural networks to make decisions. GBRL uses "gradient boosting trees" instead, which are essentially a series of simple decision rules chained together. It is built specifically for reinforcement learning, which means it can handle the unique challenges of that field, like the fact that the environment keeps changing and there is no simple right answer to learn from. A team training an AI agent might use this if they want a more transparent model, since tree-based approaches are generally easier to interpret than neural networks. The project includes ready-to-use integrations with Stable Baselines3, a popular reinforcement learning library. It also comes with implementations of common algorithms like PPO and A2C. The library is written in C++ and CUDA for speed, but you install it as a Python package. You can choose between a CPU-only version or a GPU-accelerated version if you have the right hardware. One notable design choice is that it uses a shared tree structure for both the policy, which decides what actions to take, and the value function, which estimates how good a particular state is. This sharing reduces memory and computation overhead, making it easier to scale. The project is open source under the MIT license and NVIDIA requires contributors to sign a contributor license agreement.