RLlib works with several different types of environments, including OpenAI Gym, user-defined, multi-agent, and also batched environments.
Tip
Not all environments work with all algorithms. Check out the algorithm overview for more information.
You can pass either a string name or a Python class to specify an environment. By default, strings will be interpreted as a gym environment name. Custom env classes passed directly to the trainer must take a single env_config parameter in their constructor:
import gym, ray
from ray.rllib.agents import ppo
class MyEnv(gym.Env):
def __init__(self, env_config):
self.action_space = <gym.Space>
self.observation_space = <gym.Space>
def reset(self):
return <obs>
def step(self, action):
return <obs>, <reward: float>, <done: bool>, <info: dict>
ray.init()
trainer = ppo.PPOTrainer(env=MyEnv, config={
"env_config": {}, # config to pass to env class
})
while True:
print(trainer.train())You can also register a custom env creator function with a string name. This function must take a single env_config (dict) parameter and return an env instance:
from ray.tune.registry import register_env
def env_creator(env_config):
return MyEnv(...) # return an env instance
register_env("my_env", env_creator)
trainer = ppo.PPOTrainer(env="my_env")For a full runnable code example using the custom environment API, see custom_env.py.
Warning
The gym registry is not compatible with Ray. Instead, always use the registration flows documented above to ensure Ray workers can access the environment.
In the above example, note that the env_creator function takes in an env_config object. This is a dict containing options passed in through your trainer. You can also access env_config.worker_index and env_config.vector_index to get the worker id and env id within the worker (if num_envs_per_worker > 0). This can be useful if you want to train over an ensemble of different environments, for example:
class MultiEnv(gym.Env):
def __init__(self, env_config):
# pick actual env based on worker and env indexes
self.env = gym.make(
choose_env_for(env_config.worker_index, env_config.vector_index))
self.action_space = self.env.action_space
self.observation_space = self.env.observation_space
def reset(self):
return self.env.reset()
def step(self, action):
return self.env.step(action)
register_env("multienv", lambda config: MultiEnv(config))Tip
When using logging in an environment, the logging configuration needs to be done inside the environment, which runs inside Ray workers. Any configurations outside the environment, e.g., before starting Ray will be ignored.
RLlib uses Gym as its environment interface for single-agent training. For more information on how to implement a custom Gym environment, see the gym.Env class definition. You may find the SimpleCorridor example useful as a reference.
Tip
Also check out the scaling guide for RLlib training.
There are two ways to scale experience collection with Gym environments:
- Vectorization within a single process: Though many envs can achieve high frame rates per core, their throughput is limited in practice by policy evaluation between steps. For example, even small TensorFlow models incur a couple milliseconds of latency to evaluate. This can be worked around by creating multiple envs per process and batching policy evaluations across these envs.
You can configure{"num_envs_per_worker": M}to have RLlib createMconcurrent environments per worker. RLlib auto-vectorizes Gym environments via VectorEnv.wrap().
- Distribute across multiple processes: You can also have RLlib create multiple processes (Ray actors) for experience collection. In most algorithms this can be controlled by setting the
{"num_workers": N}config.
You can also combine vectorization and distributed execution, as shown in the above figure. Here we plot just the throughput of RLlib policy evaluation from 1 to 128 CPUs. PongNoFrameskip-v4 on GPU scales from 2.4k to ∼200k actions/s, and Pendulum-v0 on CPU from 15k to 1.5M actions/s. One machine was used for 1-16 workers, and a Ray cluster of four machines for 32-128 workers. Each worker was configured with num_envs_per_worker=64.
Some environments may be very resource-intensive to create. RLlib will create num_workers + 1 copies of the environment since one copy is needed for the driver process. To avoid paying the extra overhead of the driver copy, which is needed to access the env's action and observation spaces, you can defer environment initialization until reset() is called.
RLlib will auto-vectorize Gym envs for batch evaluation if the num_envs_per_worker config is set, or you can define a custom environment class that subclasses VectorEnv to implement vector_step() and vector_reset().
Note that auto-vectorization only applies to policy inference by default. This means that policy inference will be batched, but your envs will still be stepped one at a time. If you would like your envs to be stepped in parallel, you can set "remote_worker_envs": True. This will create env instances in Ray actors and step them in parallel. These remote processes introduce communication overheads, so this only helps if your env is very expensive to step / reset.
When using remote envs, you can control the batching level for inference with remote_env_batch_wait_ms. The default value of 0ms means envs execute asynchronously and inference is only batched opportunistically. Setting the timeout to a large value will result in fully batched inference and effectively synchronous environment stepping. The optimal value depends on your environment step / reset time, and model inference speed.
A multi-agent environment is one which has multiple acting entities per step, e.g., in a traffic simulation, there may be multiple "car"- and "traffic light" agents in the environment. The model for multi-agent in RLlib is as follows: (1) as a user, you define the number of policies available up front, and (2) a function that maps agent ids to policy ids. This is summarized by the below figure:
The environment itself must subclass the MultiAgentEnv interface, which can return observations and rewards from multiple ready agents per step:
# Example: using a multi-agent env
> env = MultiAgentTrafficEnv(num_cars=20, num_traffic_lights=5)
# Observations are a dict mapping agent names to their obs. Only those
# agents' names that require actions in the next call to `step()` will
# be present in the returned observation dict.
> print(env.reset())
{
"car_1": [[...]],
"car_2": [[...]],
"traffic_light_1": [[...]],
}
# In the following call to `step`, actions should be provided for each
# agent that returned an observation before:
> new_obs, rewards, dones, infos = env.step(actions={"car_1": ..., "car_2": ..., "traffic_light_1": ...})
# Similarly, new_obs, rewards, dones, etc. also become dicts
> print(rewards)
{"car_1": 3, "car_2": -1, "traffic_light_1": 0}
# Individual agents can early exit; The entire episode is done when "__all__" = True
> print(dones)
{"car_2": True, "__all__": False}If all the agents will be using the same algorithm class to train, then you can setup multi-agent training as follows:
trainer = pg.PGAgent(env="my_multiagent_env", config={
"multiagent": {
"policies": {
# the first tuple value is None -> uses default policy
"car1": (None, car_obs_space, car_act_space, {"gamma": 0.85}),
"car2": (None, car_obs_space, car_act_space, {"gamma": 0.99}),
"traffic_light": (None, tl_obs_space, tl_act_space, {}),
},
"policy_mapping_fn":
lambda agent_id:
"traffic_light" # Traffic lights are always controlled by this policy
if agent_id.startswith("traffic_light_")
else random.choice(["car1", "car2"]) # Randomly choose from car policies
},
})
while True:
print(trainer.train())RLlib will create three distinct policies and route agent decisions to its bound policy. When an agent first appears in the env, policy_mapping_fn will be called to determine which policy it is bound to. RLlib reports separate training statistics for each policy in the return from train(), along with the combined reward.
Here is a simple example training script in which you can vary the number of agents and policies in the environment. For how to use multiple training methods at once (here DQN and PPO), see the two-trainer example. Metrics are reported for each policy separately, for example:
Result for PPO_multi_cartpole_0:
episode_len_mean: 34.025862068965516
episode_reward_max: 159.0
episode_reward_mean: 86.06896551724138
info:
policy_0:
cur_lr: 4.999999873689376e-05
entropy: 0.6833480000495911
kl: 0.010264254175126553
policy_loss: -11.95590591430664
total_loss: 197.7039794921875
vf_explained_var: 0.0010995268821716309
vf_loss: 209.6578826904297
policy_1:
cur_lr: 4.999999873689376e-05
entropy: 0.6827034950256348
kl: 0.01119876280426979
policy_loss: -8.787769317626953
total_loss: 88.26161193847656
vf_explained_var: 0.0005457401275634766
vf_loss: 97.0471420288086
policy_reward_mean:
policy_0: 21.194444444444443
policy_1: 21.798387096774192To scale to hundreds of agents, MultiAgentEnv batches policy evaluations across multiple agents internally. It can also be auto-vectorized by setting num_envs_per_worker > 1.
PettingZoo is a repository of over 50 diverse multi-agent environments. However, the API is not directly compatible with rllib, but it can be converted into an rllib MultiAgentEnv like in this example
from ray.tune.registry import register_env
# import the pettingzoo environment
from pettingzoo.butterfly import prison_v2
# import rllib pettingzoo interface
from ray.rllib.env import PettingZooEnv
# define how to make the environment. This way takes an optional environment config, num_floors
env_creator = lambda config: prison_v2.env(num_floors=config.get("num_floors", 4))
# register that way to make the environment under an rllib name
register_env('prison', lambda config: PettingZooEnv(env_creator(config)))
# now you can use `prison` as an environment
# you can pass arguments to the environment creator with the env_config option in the config
config['env_config'] = {"num_floors": 5}A more complete example is here: pettingzoo_env.py
The rock_paper_scissors_multiagent.py example demonstrates several types of policies competing against each other: heuristic policies of repeating the same move, beating the last opponent move, and learned LSTM and feedforward policies.
TensorBoard output of running the rock-paper-scissors example, where a learned policy faces off between a random selection of the same-move and beat-last-move heuristics. Here the performance of heuristic policies vs the learned policy is compared with LSTM enabled (blue) and a plain feed-forward policy (red). While the feedforward policy can easily beat the same-move heuristic by simply avoiding the last move taken, it takes a LSTM policy to distinguish between and consistently beat both policies.
Note
With ModelV2, you can put layers in global variables and straightforwardly share those layer objects between models instead of using variable scopes.
RLlib will create each policy's model in a separate tf.variable_scope. However, variables can still be shared between policies by explicitly entering a globally shared variable scope with tf.VariableScope(reuse=tf.AUTO_REUSE):
with tf.variable_scope(
tf.VariableScope(tf.AUTO_REUSE, "name_of_global_shared_scope"),
reuse=tf.AUTO_REUSE,
auxiliary_name_scope=False):
<create the shared layers here>There is a full example of this in the example training script.
Here are two ways to implement a centralized critic compatible with the multi-agent API:
Strategy 1: Sharing experiences in the trajectory preprocessor:
The most general way of implementing a centralized critic involves modifying the postprocess_trajectory method of a custom policy, which has full access to the policies and observations of concurrent agents via the other_agent_batches and episode arguments. The batch of critic predictions can then be added to the postprocessed trajectory. Here's an example:
def postprocess_trajectory(policy, sample_batch, other_agent_batches, episode):
agents = ["agent_1", "agent_2", "agent_3"] # simple example of 3 agents
global_obs_batch = np.stack(
[other_agent_batches[agent_id][1]["obs"] for agent_id in agents],
axis=1)
# add the global obs and global critic value
sample_batch["global_obs"] = global_obs_batch
sample_batch["central_vf"] = self.sess.run(
self.critic_network, feed_dict={"obs": global_obs_batch})
return sample_batchTo update the critic, you'll also have to modify the loss of the policy. For an end-to-end runnable example, see examples/centralized_critic.py.
Strategy 2: Sharing observations through an observation function:
Alternatively, you can use an observation function to share observations between agents. In this strategy, each observation includes all global state, and policies use a custom model to ignore state they aren't supposed to "see" when computing actions. The advantage of this approach is that it's very simple and you don't have to change the algorithm at all -- just use the observation func (i.e., like an env wrapper) and custom model. However, it is a bit less principled in that you have to change the agent observation spaces to include training-time only information. You can find a runnable example of this strategy at examples/centralized_critic_2.py.
It is common to have groups of agents in multi-agent RL. RLlib treats agent groups like a single agent with a Tuple action and observation space. The group agent can then be assigned to a single policy for centralized execution, or to specialized multi-agent policies such as :ref:`Q-Mix <qmix>` that implement centralized training but decentralized execution. You can use the MultiAgentEnv.with_agent_groups() method to define these groups:
.. literalinclude:: ../../rllib/env/multi_agent_env.py :language: python :start-after: __grouping_doc_begin__ :end-before: __grouping_doc_end__
For environments with multiple groups, or mixtures of agent groups and individual agents, you can use grouping in conjunction with the policy mapping API described in prior sections.
Hierarchical training can sometimes be implemented as a special case of multi-agent RL. For example, consider a three-level hierarchy of policies, where a top-level policy issues high level actions that are executed at finer timescales by a mid-level and low-level policy. The following timeline shows one step of the top-level policy, which corresponds to two mid-level actions and five low-level actions:
top_level ---------------------------------------------------------------> top_level --->
mid_level_0 -------------------------------> mid_level_0 ----------------> mid_level_1 ->
low_level_0 -> low_level_0 -> low_level_0 -> low_level_1 -> low_level_1 -> low_level_2 ->
This can be implemented as a multi-agent environment with three types of agents. Each higher-level action creates a new lower-level agent instance with a new id (e.g., low_level_0, low_level_1, low_level_2 in the above example). These lower-level agents pop in existence at the start of higher-level steps, and terminate when their higher-level action ends. Their experiences are aggregated by policy, so from RLlib's perspective it's just optimizing three different types of policies. The configuration might look something like this:
"multiagent": {
"policies": {
"top_level": (custom_policy or None, ...),
"mid_level": (custom_policy or None, ...),
"low_level": (custom_policy or None, ...),
},
"policy_mapping_fn":
lambda agent_id:
"low_level" if agent_id.startswith("low_level_") else
"mid_level" if agent_id.startswith("mid_level_") else "top_level"
"policies_to_train": ["top_level"],
},In this setup, the appropriate rewards for training lower-level agents must be provided by the multi-agent env implementation. The environment class is also responsible for routing between the agents, e.g., conveying goals from higher-level agents to lower-level agents as part of the lower-level agent observation.
See this file for a runnable example: hierarchical_training.py.
In many situations, it does not make sense for an environment to be "stepped" by RLlib. For example, if a policy is to be used in a web serving system, then it is more natural for an agent to query a service that serves policy decisions, and for that service to learn from experience over time. This case also naturally arises with external simulators (e.g. Unity3D, other game engines, or the Gazebo robotics simulator) that run independently outside the control of RLlib, but may still want to leverage RLlib for training.
A Unity3D soccer game being learnt by RLlib via the ExternalEnv API.
RLlib provides the ExternalEnv class for this purpose.
Unlike other envs, ExternalEnv has its own thread of control. At any point, agents on that thread can query the current policy for decisions via self.get_action() and reports rewards, done-dicts, and infos via self.log_returns().
This can be done for multiple concurrent episodes as well.
Take a look at the examples here for a simple "CartPole-v0" server and n client(s) scripts, in which we setup an RLlib policy server that listens on one or more ports for client connections and connect several clients to this server to learn the env.
Another example shows, how to run a similar setup against a Unity3D external game engine.
ExternalEnv provides a self.log_action() call to support off-policy actions. This allows the client to make independent decisions, e.g., to compare two different policies, and for RLlib to still learn from those off-policy actions. Note that this requires the algorithm used to support learning from off-policy decisions (e.g., DQN).
.. seealso::
`Offline Datasets <rllib-offline.html>`__ provide higher-level interfaces for working with off-policy experience datasets.
For applications that are running entirely outside the Ray cluster (i.e., cannot be packaged into a Python environment of any form), RLlib provides the PolicyServerInput application connector, which can be connected to over the network using PolicyClient instances.
You can configure any Trainer to launch a policy server with the following config:
trainer_config = {
# An environment class is still required, but it doesn't need to be runnable.
# You only need to define its action and observation space attributes.
# See examples/serving/unity3d_server.py for an example using a RandomMultiAgentEnv stub.
"env": YOUR_ENV_STUB,
# Use the policy server to generate experiences.
"input": (
lambda ioctx: PolicyServerInput(ioctx, SERVER_ADDRESS, SERVER_PORT)
),
# Use the existing trainer process to run the server.
"num_workers": 0,
# Disable OPE, since the rollouts are coming from online clients.
"input_evaluation": [],
}Clients can then connect in either local or remote inference mode. In local inference mode, copies of the policy are downloaded from the server and cached on the client for a configurable period of time. This allows actions to be computed by the client without requiring a network round trip each time. In remote inference mode, each computed action requires a network call to the server.
Example:
client = PolicyClient("http://localhost:9900", inference_mode="local")
episode_id = client.start_episode()
...
action = client.get_action(episode_id, cur_obs)
...
client.end_episode(episode_id, last_obs)To understand the difference between standard envs, external envs, and connecting with a PolicyClient, refer to the following figure:
Try it yourself by launching either a simple CartPole server (see below), and connecting it to any number of clients (cartpole_client.py) or run a Unity3D learning sever against distributed Unity game engines in the cloud.
CartPole Example:
# Start the server by running:
>>> python rllib/examples/serving/cartpole_server.py --run=PPO
--
-- Starting policy server at localhost:9900
--
# To connect from a client with inference_mode="remote".
>>> python rllib/examples/serving/cartpole_client.py --inference-mode=remote
Total reward: 10.0
Total reward: 58.0
...
Total reward: 200.0
...
# To connect from a client with inference_mode="local" (faster).
>>> python rllib/examples/serving/cartpole_client.py --inference-mode=local
Querying server for new policy weights.
Generating new batch of experiences.
Total reward: 13.0
Total reward: 11.0
...
Sending batch of 1000 steps back to server.
Querying server for new policy weights.
...
Total reward: 200.0
...For the best performance, we recommend using inference_mode="local" when possible.
For more complex / high-performance environment integrations, you can instead extend the low-level BaseEnv class. This low-level API models multiple agents executing asynchronously in multiple environments. A call to BaseEnv:poll() returns observations from ready agents keyed by 1) their environment, then 2) agent ids. Actions for those agents are sent back via BaseEnv:send_actions(). BaseEnv is used to implement all the other env types in RLlib, so it offers a superset of their functionality. For example, BaseEnv is used to implement dynamic batching of observations for inference over multiple simulator actors.