dopamine

Why we included this project

Researchers and grad students who want to test a new reinforcement learning idea against established baselines will find one of the most approachable codebases to start from. It ships compact implementations of the algorithms people actually cite in papers: DQN, C51, Rainbow, IQN, and on the JAX side SAC and PPO, so you can benchmark a variation without reimplementing the classics from scratch. The flat class hierarchy is a deliberate choice, with no abstract base class, and the docs walk you through editing an agent directly or building one from scratch, which keeps the barrier to speculative research low. Configuration runs through the gin framework, letting you swap hyperparameters or whole environments (Atari, Cartpole, Acrobot, or Mujoco continuous control) via config files rather than scattering changes across code. Because everything is wired for checkpointing and reproducible evaluation following the standard Machado et al. protocol, comparing your results against published runs is straightforward. If your project is RL research rather than deploying agents into a product, this is a solid foundation.