Overview | Installation | Documentation | Agents | Examples | Paper | Blog post

Acme is a library of reinforcement learning (RL) agents and agent building blocks. Acme strives to expose simple, efficient, and readable agents, that serve both as reference implementations of popular algorithms and as strong baselines, while still providing enough flexibility to do novel research. The design of Acme also attempts to provide multiple points of entry to the RL problem at differing levels of complexity.

If you just want to get started using Acme quickly, the main thing to know about the library is that we expose a number of agent implementations and an EnvironmentLoop primitive that can be used as follows:

loop = acme.EnvironmentLoop(environment, agent)
loop.run()

This will run a simple loop in which the given agent interacts with its environment and learns from this interaction. This assumes an agent instance (implementations of which you can find here) and an environment instance which implements the DeepMind Environment API. Each individual agent also includes a README.md file describing the implementation in more detail. Of course, these two lines of code definitely simplify the picture. To actually get started, take a look at the detailed working code examples found in our examples subdirectory which show how to instantiate a few agents and environments. We also include a quickstart notebook.

Acme also tries to maintain this level of simplicity while either diving deeper into the agent algorithms or by using them in more complicated settings. An overview of Acme along with more detailed descriptions of its underlying components can be found by referring to the documentation. And we also include a tutorial notebook which describes in more detail the underlying components behind a typical Acme agent and how these can be combined to form a novel implementation.

ℹ️ Acme is first and foremost a framework for RL research written by researchers, for researchers. We use it for our own work on a daily basis. So with that in mind, while we will make every attempt to keep everything in good working order, things may break occasionally. But if so we will make our best effort to fix them as quickly as possible!

We have tested acme on Python 3.6 & 3.7.

1. Optional: We strongly recommend using a Python virtual environment to manage your dependencies in order to avoid version conflicts:

   python3 -m venv acme
   source acme/bin/activate
   pip install --upgrade pip setuptools

2. To install the core libraries (including Reverb, our storage backend):

   pip install dm-acme
   pip install dm-acme[reverb]

3. To install dependencies for our JAX- or TensorFlow-based agents:

   pip install dm-acme[tf]
   # and/or
   pip install dm-acme[jax]

4. Finally, to install a few example environments (including gym, dm_control, and bsuite):

   pip install dm-acme[envs]

If you use Acme in your work, please cite the accompanying technical report:

@article{hoffman2020acme,
    title={Acme: A Research Framework for Distributed Reinforcement Learning},
    author={Matt Hoffman and Bobak Shahriari and John Aslanides and Gabriel
        Barth-Maron and Feryal Behbahani and Tamara Norman and Abbas Abdolmaleki
        and Albin Cassirer and Fan Yang and Kate Baumli and Sarah Henderson and
        Alex Novikov and Sergio Gómez Colmenarejo and Serkan Cabi and Caglar
        Gulcehre and Tom Le Paine and Andrew Cowie and Ziyu Wang and Bilal Piot
        and Nando de Freitas},
    year={2020},
    journal={arXiv preprint arXiv:2006.00979},
    url={https://arxiv.org/abs/2006.00979},
}