What is di-engine?

DI-engine is a Python framework for building and running reinforcement learning experiments. Reinforcement learning is a branch of machine learning where a program learns to make decisions by trial and error, getting feedback in the form of rewards or penalties. DI-engine provides the building blocks needed to run these experiments without writing everything from scratch. The framework organizes experiments around three main concepts: an environment (the world the program operates in), a policy (the rules the program follows when deciding what to do), and a model (the neural network that does the actual computation). These pieces plug into a shared task-and-middleware system that handles the coordination between collecting data, training the model, and evaluating results. It ships with implementations of a large number of well-known reinforcement learning algorithms, covering the most common single-agent approaches as well as multi-agent setups, imitation learning (where the program learns by watching examples rather than trial and error), offline learning from pre-collected data, and algorithms that combine language models with reinforcement learning. The list is extensive and primarily aimed at researchers who want a reference implementation or a baseline to compare against. Beyond academic benchmarks, the repository links to several real-world application projects built on top of DI-engine, including an autonomous driving platform, a traffic light controller, a biological sequence search tool, and a StarCraft II agent. These projects show the kinds of problems the framework is meant to address. DI-engine works with both PyTorch and JAX. Documentation is available in both English and Chinese. Installation is via pip or conda. The full README is longer than what was shown.