rlstack 1.1.0

rlstack: A Minimal RL Library

rlstack is a minimal RL library that can simulate highly parallelized, infinite horizon environments, and can train a PPO policy using those environments, achieving up to 1M environment transitions (and one policy update) per second using a single NVIDIA RTX 2080.

• Documentation: https://theogognf.github.io/rlstack/

• PyPI: https://pypi.org/project/rlstack/

• Repository: https://github.com/theOGognf/rlstack

Quick Start

pip install rlstack

git clone https://github.com/theOGognf/rlstack.git
pip install ./rlstack/

from rlstack import Trainer
from rlstack.env import DiscreteDummyEnv

trainer = Trainer(DiscreteDummyEnv)
trainer.run()

from rlstack import Algorithm
from rlstack.env import DiscreteDummyEnv

algo = Algorithm(DiscreteDummyEnv)
algo.collect()
algo.step()

Concepts

rlstack is minimal in that it limits the number of interfaces required for training a policy with PPO, even for customized policies, without restrictions on observation and action specs, custom models, and custom action distributions.

rlstack is built around six key concepts:

• The environment: The simulation that the policy learns to interact with. The environment is always user-defined.

• The model: The policy parameterization that determines how the policy processes environment observations and how parameters for the action distribution are generated. The model is usually user-defined (default models are sometimes sufficient depending on the environment’s observation and action specs).

• The action distribution: The mechanism for representing actions conditioned on environment observations and model outputs. Environment actions are ultimately sampled from the action distribution. The action distribution is sometimes user-defined (default action distributions are usually sufficient depending on the environment’s observation and action specs).

• The policy: The union of the model and the action distribution that actually calls and samples from the model and action distribution, respectively. The policy handles some pre/post -processing on its I/O to make it more convenient to sample from the model and action distribution together. The policy is rarely user-defined.

• The algorithm: The PPO implementation that uses the environment to train the policy (i.e., update the model’s parameters). All hyperparameters and customizations are set with the algorithm. The algorithm is rarely user-defined.

• The trainer: The high-level interface for using the algorithm to train indefinitely or until some condition is met. The trainer directly integrates with MLflow to track experiments and training progress. The trainer is never user-defined.

Quick Examples

from rlstack import SquashedNormal, Trainer
from rlstack.env import ContinuousDummyEnv

trainer = Trainer(
    ContinuousDummyEnv,
    distribution_cls=SquashedNormal,
    gae_lambda=0.99,
    gamma=0.99,
)
trainer.run()

from rlstack import RecurrentTrainer
from rlstack.env import DiscreteDummyEnv

trainer = RecurrentTrainer(DiscreteDummyEnv)
trainer.run()

from rlstack import Trainer
from rlstack.env import DiscreteDummyEnv

trainer = Trainer(DiscreteDummyEnv, device="cuda")
trainer.run()

import torch.optim as optim

from rlstack import Trainer
from rlstack.env import DiscreteDummyEnv

trainer = Trainer(
    DiscreteDummyEnv,
    optimizer_cls=optim.SGD,
    accumulate_grads=True,
    enable_amp=True,
    sgd_minibatch_size=8192,
    device="cuda",
)
trainer.run()

from rlstack import Trainer
from rlstack.conditions import Plateaus
from rlstack.env import DiscreteDummyEnv

trainer = Trainer(DiscreteDummyEnv)
trainer.run(stop_conditions=[Plateaus("returns/mean", rtol=0.05)])

Why rlstack?

TL;DR: rlstack focuses on a niche subset of RL that simplifies the overall library while allowing fast and fully customizable environments, models, and action distributions.

There are many high quality, open-sourced RL libraries. Most of them take on the daunting task of being a monolithic, one-stop-shop for everything RL, attempting to support as many algorithms, environments, models, and compute capabilities as possible. Naturely, this monolothic goal has some drawbacks:

• The software becomes more dense with each supported feature, making the library all-the-more difficult to customize for a specific use case.

• The software becomes less performant for a specific use case. RL practitioners typically end up accepting the cost of transitioning to expensive and difficult-to-manage compute clusters to get results faster.

There’s a handful of high quality, open-sourced RL libraries that tradeoff feature richness to reduce these drawbacks. However, each library still doesn’t provide enough speed benefit to warrant the switch from a monolithic repo, or is still too complex to adapt to a specific use case.

rlstack is a niche RL library that finds a goldilocks zone between the feature support and speed/complexity tradeoff by making some key assumptions:

• Environments are highly parallelized and their parallelization is entirely managed within the environment. This allows rlstack to ignore distributed computing design considerations.

• Environments are infinite horizon (i.e., they have no terminal conditions). This allows rlstack to reset environments at the same, fixed horizon intervals, greatly simplifying environment and algorithm implementations.

• The only supported ML framework is PyTorch and the only supported algorithm is PPO. This allows rlstack to ignore layers upon layers of abstraction, greatly simplifying the overall library implementation.

The end result is a minimal and high throughput library that can train policies to solve complex tasks on a single NVIDIA RTX 2080 within minutes.

Unfortunately, this means rlstack doesn’t support as many use cases as a monolithic RL library might. In fact, rlstack is probably a bad fit for your use case if:

• Your environment isn’t parallelizable.

• Your environment must contain terminal conditions and can’t be reformulated as an infinite horizon task.

• You want to use an ML framework that isn’t PyTorch or you want to use an algorithm that isn’t a variant of PPO.

However, if rlstack does fit your use case, it can do wonders for your RL workflow.

Related Projects

• RL Games: RL Games is a high performance RL library built around popular environment protocols.

• RLlib: Ray’s RLlib is the industry standard RL library that supports many popular RL algorithms. RLlib can scale RL workloads from your laptop all the way to the cloud with little-to-no changes to your code.

• Sample Factory: Sample Factory provides an efficient and high quality implementation of PPO with a focus on accelerating training for a single machine with support for a wide variety of environment protocols.

• SKRL: SKRL focuses on readability, simplicity, and transparency of RL algorithm implementations with support for a wide variety of environment protocols.

• Stable Baselines 3: Stable Baselines 3 is a set of reliable and user-friendly RL algorithm implementations that integrate with a rich set of features desirable by most practitioners and use cases.

• TorchRL: TorchRL is PyTorch’s RL library that’s focused on efficient, modular, documented, and tested RL building blocks and algorithm implementations aimed at supporting research in RL. TorchRL is a direct dependency of rlstack.