keras-rl

Why we included this project

Teams that want to try deep reinforcement learning without coding every algorithm themselves will find keras-rl a gentle place to start. It packages several established methods, including DQN and its Double and Dueling variants, DDPG, CEM, and Deep SARSA, behind a plain Python API that plugs into Keras models. Since it connects to OpenAI Gym out of the box, a few lines of code are enough to watch an agent learn on CartPole, Pendulum, or Atari. The abstract base classes for agents and environments make it easy to bring in your own problem, so it works well as a teaching and prototyping tool. Treat it as a way to see how RL algorithms behave rather than as infrastructure for heavy production workloads.