ai-traineree 0.5.1

Yet another zoo of (Deep) Reinforcment Learning methods in Python using PyTorch

ai-traineree

The intention is to have a zoo of Deep Reinforcment Learning methods and showcasing their application on some environments.

Read more in the doc: ReadTheDocs AI-Traineree.

Why another?

The main reason is the implemention philosophy. We strongly believe that agents should be emerged in the environment and not the other way round. Majority of the popular implementations pass environment instance to the agent as if the agent was the focus point. This might ease implementation of some algorithms but it isn't representative of the world; agents want to control the environment but that doesn't mean they can/should.

That, and using PyTorch instead of Tensorflow or JAX.

Quick start

To get started with training your RL agent you need three things: an agent, an environment and a runner. Let's say you want to train a DQN agent on OpenAI CartPole-v1:

from ai_traineree.agents.dqn import DQNAgent
from ai_traineree.runners.env_runner import EnvRunner
from ai_traineree.tasks import GymTask

task = GymTask('CartPole-v1')
agent = DQNAgent(task.obs_space, task.action_space)
env_runner = EnvRunner(task, agent)

scores = env_runner.run()

or execute one of provided examples

$ python -m examples.cart_dqn

That's it.

Installation

PyPi (recommended)

The quickest way to install package is through pip.

$ pip install ai-traineree

Git repository clone

As usual with Python, the expectation is to have own virtual environment and then pip install requirements. For example,

> python -m venv .venv
> git clone [e-mail]:laszukdawid/ai-traineree.git
> source .venv/bin/activate
> python setup.py install

Current state

Playing gym

One way to improve learning speed is to simply show them how to play or, more researchy/creepy, provide a proper seed. This isn't a general rule, since some algorithms train better without any human interaction, but since you're on GitHub... that's unlikely your case. Currently there's a script interact.py which uses OpenAI Gym's play API to record moves and AI Traineree to store them in a buffer. Such buffers can be loaded by agents on initiation.

This is just a beginning and there will be more work on these interactions.

Requirement: Install pygame.

Agents

Short	Progress	Link	Full name	Doc
DQN	Implemented	DeepMind	Deep Q-learning Network	Doc
DDPG	Implemented	arXiv	Deep Deterministic Policy Gradient	Doc
D4PG	Implemented	arXiv	Distributed Distributional Deterministic Policy Gradients	Doc
TD3	Implemented	arXiv	Twine Delayed Deep Deterministic policy gradient	Doc
PPO	Implemented	arXiv	Proximal Policy Optimization	Doc
SAC	Implemented	arXiv	Soft Actor Critic	Doc
TRPO		arXiv	Trust Region Policy Optimization
RAINBOW	Implemented	arXiv	DQN with a few improvements	Doc

Multi agents

We provide both Multi Agents agents entities and means to execute them against supported (below) environements. However, that doesn't mean one can be used without the other.

Short	Progress	Link	Full name	Doc
IQL	Implemented		Independent Q-Learners	Doc
MADDPG	Implemented	arXiv	Multi agent DDPG	Doc

Loggers

Supports using Tensorboard (via PyTorch's SummaryWriter) and Neptune to display metrics. Wrappers are provided as TensorboardLogger and NeptuneLogger.

Note: In order to use Neptune one needs to install neptune-client (pip install neptune-client).

Environments

Name	Progress	Link
OpenAI Gym - Classic	Done
OpenAI Gym - Atari	Done
OpenAI Gym - MuJoCo	Not interested.
PettingZoo	Initial support	Page / GitHub
Unity ML	Somehow supported.	Page
MAME Linux emulator	Interested.	Official page

Development

We are open to any contributions. If you want to contribute but don't know what then feel free to reach out (see Contact below). The best way to start is through updating documentation and adding tutorials. In addition there are many other things that we know of which need improvement but also plenty that we don't know of.

Setting up development environment requires installing dev and test extra packages. The dev extras are for mainly for linting and formatting, and the test is for running tests. We recommend using pip so to install everything requires for development run

$ pip install -e .[dev,test]

Once installed, please configure your IDE to use black as formatter, pycodestyle as linter, and isort for sorting imports. All these are included in the dev extra packages.

Contact

Should we focus on something specificallly? Let us know by opening a feature request GitHub issue or contacting through [e-mail].

Citing project

@misc{ai-traineree,
  author = {Laszuk, Dawid},
  title = {AI Traineree: Reinforcement learning toolset},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/laszukdawid/ai-traineree}},
}