.. include:: /_includes/rllib/we_are_hiring.rst .. sphinx_rllib_readme_begin .. _rllib-index: RLlib: Industry-Grade, Scalable Reinforcement Learning ====================================================== .. include:: /_includes/rllib/new_api_stack.rst .. image:: images/rllib-logo.png :align: center .. sphinx_rllib_readme_end .. todo (sven): redo toctree: suggestion: getting-started key-concepts rllib-env (single-agent) ... <- multi-agent ... <- external ... <- hierarchical algorithm-configs rllib-algorithms (overview of all available algos) dev-guide (replaces user-guides) debugging scaling-guide fault-tolerance checkpoints callbacks metrics-logger rllib-advanced-api algorithm (general description of how algos work) rl-modules rllib-offline single-agent-episode multi-agent-episode connector-v2 rllib-learner env-runners rllib-examples new-api-stack-migration-guide package_ref/index .. toctree:: :hidden: getting-started key-concepts rllib-env algorithm-config rllib-algorithms user-guides rllib-examples new-api-stack-migration-guide package_ref/index .. sphinx_rllib_readme_2_begin **RLlib** is an open source library for reinforcement learning (**RL**), offering support for production-level, highly scalable, and fault-tolerant RL workloads, while maintaining simple and unified APIs for a large variety of industry applications. Whether training policies in a **multi-agent** setup, from historic **offline** data, or using **externally connected simulators**, RLlib offers simple solutions for each of these autonomous decision making needs and enables you to start running your experiments within hours. Industry leaders use RLlib in production in many different verticals, such as `gaming `_, `robotics `_, `finance `_, `climate- and industrial control `_, `manufacturing and logistics `_, `automobile `_, and `boat design `_. RLlib in 60 seconds ------------------- .. figure:: images/rllib-index-header.svg It only takes a few steps to get your first RLlib workload up and running on your laptop. Install RLlib and `PyTorch `__, as shown below: .. code-block:: bash pip install "ray[rllib]" torch .. note:: For installation on computers running Apple Silicon, such as M1, `follow instructions here. `_ .. note:: To be able to run the Atari or MuJoCo examples, you also need to do: .. code-block:: bash pip install "gymnasium[atari,accept-rom-license,mujoco]" This is all, you can now start coding against RLlib. Here is an example for running the :ref:`PPO Algorithm ` on the `Taxi domain `__. You first create a `config` for the algorithm, which defines the :ref:`RL environment ` and any other needed settings and parameters. .. testcode:: from ray.rllib.algorithms.ppo import PPOConfig from ray.rllib.connectors.env_to_module import FlattenObservations # Configure the algorithm. config = ( PPOConfig() .environment("Taxi-v3") .env_runners( num_env_runners=2, # Observations are discrete (ints) -> We need to flatten (one-hot) them. env_to_module_connector=lambda env: FlattenObservations(), ) .evaluation(evaluation_num_env_runners=1) ) Next, ``build`` the algorithm and ``train`` it for a total of five iterations. One training iteration includes parallel, distributed sample collection by the :py:class:`~ray.rllib.env.env_runner.EnvRunner` actors, followed by loss calculation on the collected data, and a model update step. .. testcode:: from pprint import pprint # Build the algorithm. algo = config.build_algo() # Train it for 5 iterations ... for _ in range(5): pprint(algo.train()) At the end of your script, you evaluate the trained Algorithm and release all its resources: .. testcode:: # ... and evaluate it. pprint(algo.evaluate()) # Release the algo's resources (remote actors, like EnvRunners and Learners). algo.stop() You can use any `Farama-Foundation Gymnasium `__ registered environment with the ``env`` argument. In ``config.env_runners()`` you can specify - amongst many other things - the number of parallel :py:class:`~ray.rllib.env.env_runner.EnvRunner` actors to collect samples from the environment. You can also tweak the NN architecture used by tweaking RLlib's :py:class:`~ray.rllib.core.rl_module.default_model_cnofig.DefaultModelConfig`, as well as, set up a separate config for the evaluation :py:class:`~ray.rllib.env.env_runner.EnvRunner` actors through the ``config.evaluation()`` method. :ref:`See here `, if you want to learn more about the RLlib training APIs. Also, `see here `__ for a simple example on how to write an action inference loop after training. If you want to get a quick preview of which **algorithms** and **environments** RLlib supports, click the dropdowns below: .. dropdown:: **RLlib Algorithms** :animate: fade-in-slide-down +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | **On-Policy** | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`PPO (Proximal Policy Optimization) ` | |single_agent| | |multi_agent| | |discr_act| | |cont_act| | |multi_gpu| | |multi_node_multi_gpu| | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | **Off-Policy** | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`SAC (Soft Actor Critic) ` | |single_agent| | |multi_agent| | | |cont_act| | |multi_gpu| | |multi_node_multi_gpu| | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`DQN/Rainbow (Deep Q Networks) ` | |single_agent| | |multi_agent| | |discr_act| | | |multi_gpu| | |multi_node_multi_gpu| | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | **High-throughput Architectures** | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`APPO (Asynchronous Proximal Policy Optimization) ` | |single_agent| | |multi_agent| | |discr_act| | |cont_act| | |multi_gpu| | |multi_node_multi_gpu| | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`IMPALA (Importance Weighted Actor-Learner Architecture) ` | |single_agent| | |multi_agent| | |discr_act| | | |multi_gpu| | |multi_node_multi_gpu| | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | **Model-based RL** | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`DreamerV3 ` | |single_agent| | | |discr_act| | |cont_act| | |multi_gpu| | |multi_node_multi_gpu| | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | **Offline RL and Imitation Learning** | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`BC (Behavior Cloning) ` | |single_agent| | | |discr_act| | |cont_act| | | | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`CQL (Conservative Q-Learning) ` | |single_agent| | | | |cont_act| | | | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ | :ref:`MARWIL (Advantage Re-Weighted Imitation Learning) ` | |single_agent| | | |discr_act| | |cont_act| | | | +-------------------------------------------------------------------------+----------------+---------------+-------------+------------+-------------+------------------------+ .. dropdown:: **RLlib Environments** :animate: fade-in-slide-down +-------------------------------------------------------------------------------------------+ | **Farama-Foundation Environments** | +-------------------------------------------------------------------------------------------+ | `gymnasium `__ |single_agent| | | | | .. code-block:: bash | | | | pip install "gymnasium[atari,accept-rom-license,mujoco]"`` | | | | .. code-block:: python | | | | config.environment("CartPole-v1") # Classic Control | | config.environment("ale_py:ALE/Pong-v5") # Atari | | config.environment("Hopper-v5") # MuJoCo | +-------------------------------------------------------------------------------------------+ | `PettingZoo `__ |multi_agent| | | | | .. code-block:: bash | | | | pip install "pettingzoo[all]" | | | | .. code-block:: python | | | | from ray.tune.registry import register_env | | from ray.rllib.env.wrappers.pettingzoo_env import PettingZooEnv | | from pettingzoo.sisl import waterworld_v4 | | register_env("env", lambda _: PettingZooEnv(waterworld_v4.env())) | | config.environment("env") | +-------------------------------------------------------------------------------------------+ | **RLlib Multi-Agent** | +-------------------------------------------------------------------------------------------+ | `RLlib's MultiAgentEnv API `__ |multi_agent| | | | | .. code-block:: python | | | | from ray.rllib.examples.envs.classes.multi_agent import MultiAgentCartPole | | from ray import tune | | tune.register_env("env", lambda cfg: MultiAgentCartPole(cfg)) | | config.environment("env", env_config={"num_agents": 2}) | | config.multi_agent( | | policies={"p0", "p1"}, | | policy_mapping_fn=lambda aid, *a, **kw: f"p{aid}", | | ) | +-------------------------------------------------------------------------------------------+ Why chose RLlib? ---------------- .. dropdown:: **Scalable and Fault-Tolerant** :animate: fade-in-slide-down RLlib workloads scale along various axes: - The number of :py:class:`~ray.rllib.env.env_runner.EnvRunner` actors to use. This is configurable through ``config.env_runners(num_env_runners=...)`` and allows you to scale the speed of your (simulator) data collection step. This `EnvRunner` axis is fully **fault tolerant**, meaning you can train against custom environments that are unstable or frequently stall execution and even place all your `EnvRunner` actors on spot machines. - The number of :py:class:`~ray.rllib.core.learner.Learner` actors to use for **multi-GPU training**. This is configurable through ``config.learners(num_learners=...)`` and you normally set this to the number of GPUs available (make sure you then also set ``config.learners(num_gpus_per_learner=1)``) or - if you do not have GPUs - you can use this setting for **DDP-style learning on CPUs** instead. .. dropdown:: **Multi-Agent Reinforcement Learning (MARL)** :animate: fade-in-slide-down RLlib natively supports multi-agent reinforcement learning (MARL), thereby allowing you to run in any complex configuration. - **Independent** multi-agent learning (the default): Every agent collects data for updating its own policy network, interpreting other agents as part of the environment. - **Collaborative** training: Train a team of agents that either all share the same policy (shared parameters) or in which some agents have their own policy network(s). You can also share value functions between all members of the team or some of them, as you see fit, thus allowing for global vs local objectives to be optimized. - **Adversarial** training: Have agents play against other agents in competitive environments. Use self-play, or league based self-play to train your agents to learn how to play throughout various stages of ever increasing difficulty. - **Any combination of the above!** Yes, you can train teams of arbitrary sizes of agents playing against other teams where the agents in each team might have individual sub-objectives and there are groups of neutral agents not participating in any competition. .. dropdown:: **Offline RL and Behavior Cloning** :animate: fade-in-slide-down **Ray.Data** has been integrated into RLlib, enabling **large-scale data ingestion** for offline RL and behavior cloning (BC) workloads. See here for a basic `tuned example for the behavior cloning algo `__ and here for how to `pre-train a policy with BC, then finetuning it with online PPO `__. .. dropdown:: **Support for External Env Clients** :animate: fade-in-slide-down **Support for externally connecting RL environments** is achieved through customizing the :py:class:`~ray.rllib.env.env_runner.EnvRunner` logic from RLlib-owned, internal gymnasium envs to external, TCP-connected Envs that act independently and may even perform their own action inference, e.g. through ONNX. See here for an example of `RLlib acting as a server with connecting external env TCP-clients `__. Learn More ---------- .. grid:: 1 2 3 3 :gutter: 1 :class-container: container pb-4 .. grid-item-card:: **RLlib Key Concepts** ^^^ Learn more about the core concepts of RLlib, such as Algorithms, environments, models, and learners. +++ .. button-ref:: rllib-key-concepts :color: primary :outline: :expand: Key Concepts .. grid-item-card:: **RL Environments** ^^^ Get started with environments supported by RLlib, such as Farama foundation's Gymnasium, Petting Zoo, and many custom formats for vectorized and multi-agent environments. +++ .. button-ref:: rllib-environments-doc :color: primary :outline: :expand: Environments .. grid-item-card:: **Models (RLModule)** ^^^ Learn how to configure RLlib's default models and implement your own custom models through the RLModule APIs, which support arbitrary architectures with PyTorch, complex multi-model setups, and multi-agent models with components shared between agents. +++ .. button-ref:: rlmodule-guide :color: primary :outline: :expand: Models (RLModule) .. grid-item-card:: **Algorithms** ^^^ See the many available RL algorithms of RLlib for on-policy and off-policy training, offline- and model-based RL, multi-agent RL, and more. +++ .. button-ref:: rllib-algorithms-doc :color: primary :outline: :expand: Algorithms Customizing RLlib ----------------- RLlib provides powerful, yet easy to use APIs for customizing all aspects of your experimental- and production training-workflows. For example, you may code your own `environments `__ in python using the `Farama Foundation's gymnasium `__ or DeepMind's OpenSpiel, provide custom `PyTorch models `_, write your own `optimizer setups and loss definitions `__, or define custom `exploratory behavior `_. .. figure:: images/rllib-new-api-stack-simple.svg :align: left :width: 850 **RLlib's API stack:** Built on top of Ray, RLlib offers off-the-shelf, distributed and fault-tolerant algorithms and loss functions, PyTorch default models, multi-GPU training, and multi-agent support. Users customize their experiments by subclassing the existing abstractions. .. sphinx_rllib_readme_2_end .. sphinx_rllib_readme_3_begin Citing RLlib ------------ If RLlib helps with your academic research, the Ray RLlib team encourages you to cite these papers: .. code-block:: @inproceedings{liang2021rllib, title={{RLlib} Flow: Distributed Reinforcement Learning is a Dataflow Problem}, author={ Wu, Zhanghao and Liang, Eric and Luo, Michael and Mika, Sven and Gonzalez, Joseph E. and Stoica, Ion }, booktitle={Conference on Neural Information Processing Systems ({NeurIPS})}, year={2021}, url={https://proceedings.neurips.cc/paper/2021/file/2bce32ed409f5ebcee2a7b417ad9beed-Paper.pdf} } @inproceedings{liang2018rllib, title={{RLlib}: Abstractions for Distributed Reinforcement Learning}, author={ Eric Liang and Richard Liaw and Robert Nishihara and Philipp Moritz and Roy Fox and Ken Goldberg and Joseph E. Gonzalez and Michael I. Jordan and Ion Stoica, }, booktitle = {International Conference on Machine Learning ({ICML})}, year={2018}, url={https://arxiv.org/pdf/1712.09381} } .. sphinx_rllib_readme_3_end .. sigils used on this page .. |single_agent| image:: /rllib/images/sigils/single-agent.svg :class: inline-figure :width: 72 .. |multi_agent| image:: /rllib/images/sigils/multi-agent.svg :class: inline-figure :width: 72 .. |discr_act| image:: /rllib/images/sigils/discr-actions.svg :class: inline-figure :width: 72 .. |cont_act| image:: /rllib/images/sigils/cont-actions.svg :class: inline-figure :width: 72 .. |multi_gpu| image:: /rllib/images/sigils/multi-gpu.svg :class: inline-figure :width: 72 .. |multi_node_multi_gpu| image:: /rllib/images/sigils/multi-node-multi-gpu.svg :class: inline-figure :alt: Only on the Anyscale Platform! :width: 72