mobile-env is an open, minimalist environment for training and evaluating coordination algorithms in wireless mobile networks. The environment allows modeling users moving around an area and can connect to one or multiple base stations. Using the Gymnasium (previously Gym) interface, the environment can be used with any reinforcement learning framework (e.g., stable-baselines or Ray RLlib) or any custom (even non-RL) coordination approach. The environment is highly configurable and can be easily extended (e.g., regarding users, movement patterns, channel models, etc.).

mobile-env supports multi-agent and centralized reinforcement learning policies. It provides various choices for rewards and observations. mobile-env is also easily extendable, so that anyone may add another channel models (e.g. path loss), movement patterns, utility functions, etc.

As an example, mobile-env can be used to study multi-cell selection in coordinated multipoint. Here, it must be decided what connections should be established among user equipments (UEs) and base stations (BSs) in order to maximize Quality of Experience (QoE) globally. To maximize the QoE of single UEs, the UE intends to connect to as many BSs as possible, which yields higher (macro) data rates. However, BSs multiplex resources among connected UEs (e.g. schedule physical resource blocks) and, therefore, UEs compete for limited resources (conflicting goals). To maximize QoE globally, the policy must recognize that (1) the data rate of any connection is governed by the channel (e.g. SNR) between UE and BS and (2) QoE of single UEs not necessarily grows linearly with increasing data rate.

^{Base station icon by Clea Doltz from the Noun Project}

Try mobile-env:

• Part I: Customizing mobile-env and single-agent RL with stable-baselines3:
• Part II: Multi-agent RL on mobile-env with Ray RLlib:

Documentation and API: ReadTheDocs

If you use mobile-env in your work, please cite our paper (author PDF):

@inproceedings{schneider2022mobileenv,
  author = {Schneider, Stefan and Werner, Stefan and Khalili, Ramin and Hecker, Artur and Karl, Holger},
  title = {mobile-env: An Open Platform for Reinforcement Learning in Wireless Mobile Networks},
  booktitle={Network Operations and Management Symposium (NOMS)},
  year = {2022},
  publisher = {IEEE/IFIP},
}

mobile-env is based on the underlying environment using in DeepCoMP, which is a combination of reinforcement learning approaches for dynamic multi-cell selection. mobile-env provides this underlying environment as open, stand-alone environment.

The simplest option is to install the latest release of mobile-env from PyPI using pip:

pip install mobile-env

This is recommended for most users. mobile-env is tested on Ubuntu, Windows, and MacOS.

Alternatively, for development, you can clone mobile-env from GitHub and install it from source. After cloning, install in "editable" mode (-e):

pip install -e .

This is equivalent to running pip install -r requirements.txt.

If you want to run tests or examples, also install the requirements in tests. For dependencies for building docs, install the requirements in docs.

import gymnasium
import mobile_env

env = gymnasium.make("mobile-medium-central-v0")
obs, info = env.reset()
done = False

while not done:
    action = ... # Your agent code here
    obs, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated
    env.render()

mobile-env supports custom channel models, movement patterns, arrival & departure models, resource multiplexing schemes and utility functions. For example, replacing the default Okumura–Hata channel model by a (simplified) path loss model can be as easy as this:

import gymnasium
import numpy as np
from mobile_env.core.base import MComCore
from mobile_env.core.channel import Channel


class PathLoss(Channel):
    def __init__(self, gamma, **kwargs):
        super().__init__(**kwargs)
        # path loss exponent
        self.gamma = gamma

    def power_loss(self, bs, ue):
        """Computes power loss between BS and UE."""
        dist = bs.point.distance(ue.point)
        loss = 10 * self.gamma * np.log10(4 * np.pi * dist * bs.frequency)
        return loss


# replace default channel model in configuration
config = MComCore.default_config()
config['channel'] = PathLoss

# pass init parameters to custom channel class!
config['channel_params'].update({'gamma': 2.0})

# create environment with custom channel model
env = gymnasium.make('mobile-small-central-v0', config=config)
# ...

If you are using movile-env, please let us know and we are happy to link to your project from the readme. You can also open a pull request yourself.

• Mohammadreza Kouchaki and Vuk Marojevic, "Actor-Critic Network for O-RAN Resource Allocation: xApp Design, Deployment, and Analysis", 2022
• Stefan Schneider, Ramin Khalili, Artur Hecker, Holger Karl, "DeepCoMP: Self-Learning Dynamic Multi-Cell Selection for Coordinated Multipoint (CoMP)", 2021

Development: @stefanbschneider and @stwerner97

We happy if you find mobile-env useful. If you have feedback or want to report bugs, feel free to open an issue. Also, we are happy to link to your projects if you use mobile-env.

We also welcome contributions: Whether you implement a new channel model, fix a bug, or just make a minor addition elsewhere, feel free to open a pull request!