Implementing a reinforcement learning environment and algorithms for networking from scratch is a difficult task. Inspired by the work of ns3-gym, we developed RL4Net (Reinforcement Learning for Networking) to facilitate the research and simulator of reinforcement learning for networking.

Below figure shows the architecture of RL4Net:

RL4Net is composed of two functional blocks:

• Environment: Environment is built on widely used ns3 network simulator ns3. We extend ns3 with six components:
  • Metric Extractor for computing quality metrics like delay and loss from ns3;
  • Computers for translating quality metrics to DRL state and reward;
  • Action Operator to get action commands from agent;
  • Action Executor for perform ns3 operations by actions;
  • ns3Env for transforming the ns3 object into DRL environment;
  • envInterface to translate between ns3 data and DRL factors.
• Agent: Agent is container of a DRL-based cognitive routing algorithm. A agent can built on various deep learning frameworks like pyTorch and Tensorflow.

• ./ns3-addon: Files to be copied into ns3 source file folder for extension. It includes:
  • ns3-src/action-executor: code for Action Executor
  • ns3-src/metric-extractor: code for Metric Extractor
  • rapidjson: an open source JSON parser and generator
  • ns3-scratch: several examples of experiments on RL4Net
• ./ns3-env: File for ns3Env block. It cinludes:
  • env-interface: code for envInterface
  • ns3-python-connector: code for connecting python and ns3 c++
• ./RL4Net-lib: Libaray files developed by us
• ./TE-trainer: Files for traning agents
• ./RLAgent: Files of agents

Attention: Switch to user root before installation is strongly recommended.

First of all, you should download RL4Net from Github:

git clone https://github.com/bupt-ipcr/RL4Net

by default, source code will be download into ./RL4Net folder.

To combine python and ns-3, RL4Net requires ZMQ and libprotoc. You can install as follow:

# to install protobuf-3.6 on ubuntu 16.04:
sudo add-apt-repository ppa:maarten-fonville/protobuf
sudo apt-get update
apt-get install libzmq5 libzmq5-dev
apt-get install libprotobuf-dev
apt-get install protobuf-compiler

Notice: libprotoc version is 3.6 if you download this way. The c++ file compiled will change at after version 3.7. Potential error will explain later.

Here we have requirements for python:

• To use RL4Net-lib, you need python>=3.6 for f-string.

• To test the project, module pytest is required you can install pytest by:

  pip install pytest -U

• (Optional) create conda enviornment You can use conda help you to manage above by:

  conda create -n NAME python=VERSION

  which NAME means your environment name and valid VERSION for this project is 3.6 and 3.7.

• (Required) Install pytorch and cuda using appropriate instructions
  You can checkout the install command in https://pytorch.org/
  It will help you manage pytorch and cuda.
  The recommended pytorch version is 1.4.0.

Since RL4Net is based on ns-3, you need to install ns-3 before use RL4Net.
The introcuction of ns-3 and how to install can be find at the official website of ns-3.
As a recommendation, you can:

1. Install dependencies
   You can install ns-3 dependencies by following official guide.
2. Use git to install ns-3-dev
   see: downloading-ns-3-using-git
   you can also install a specific version od ns-3, such as ns-3.30, but we prefer ns-3-dev.

Another possible guide is wiki of ns-3, see: wiki of installation

Now suppose you have successfuly installed ns-3-dev, you can start to install RL4Net.

As recommendation, deactivate your conda env before run setup script.
Conda may install libprotoc in your virtual environment with version >= 3.6, while your system libprotoc version is 3.6.
Thus when configure, protoc head file will read version >= 3.6 and when build, protoc head file will read libprotoc v3.6 and cause error.

python ns3_setup.py --wafdir=YOUR_WAFPATH

the YOUR_WAFPATH is correspond to the introduction of ns-3 installation, where you can execute ./waf build, typically ns-3-allinone/ns-3-dev. Remember to use absolute path.

The default value of wafdir is /ns-3-dev (notice it is subdir of '/'). As an alternative, you can copy the folder into /ns-3-dev, then run

python ns3_setup.py

pyns3 is the python module that connect python and ns3. Use pip(or pip3) to install this module with your python env(maybe conda).

pip install ns3-env/ns3-python-connector

wjwgym is a lab that helps build reinforcement learning algorithms. See: Github
Install it with pip and your python env:

pip install RL4Net-lib/wjwgym-home

The lab need numpy, torch and tensorboard. You can pre-install them, especially pytorch, by which you can choose pip/conda.

The test file can help you check if you have every in place. You can run test by:

pytest

inside folder RL4Net, it will automatically detect test files.
You can also specify file like:

pytest test_installation.py

Jun Liu ([email protected]), Beijing University of Posts and Telecommunications, China