World-Model-Based Adaptive Network Slicing

We introduce a world-model-based RL framework for adaptive network slicing, leveraging DreamerV3 based world model.

Installation

Install Conda

First, install Conda by following the instructions on the Miniconda website. You can skip this part if you have already installed Conda.

mkdir -p ~/miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
rm -rf ~/miniconda3/miniconda.sh

Then initialize your Miniconda on bash shells:

~/miniconda3/bin/conda init bash
source ~/.bashrc

Create a Conda Environment

Create a new Conda environment by running:

conda create -n slice python=3.8 -y

Activate the environment by running:

conda activate slice

Install the environment for WM-based network slicing.

pip install -r requirements.txt

Training and Evaluation

The environment is located at the embodied\envs\slice_env.py. You can change the system dynamic and reward function by adjusting the parameter in the script.

To train the World Model, run the script

python train.py

To visualize the training reward, you can use tensorboard to monitor the training process:

tensorboard --logdir ./logdir

To evaluate the World Model, specify the key from_checkpoint in the configuration file *dreamerv3/configs.yaml* by the path of the checkpoint to be evaluated. For example:

from_checkpoint: '/home/sliceWM/logdir/108-slice/checkpoint.ckpt'

Then run evaluation script by:

python eval.py

The evaluation script records all the metrics in a csv file recorded_metrics.csv. To calculate the metrics in the evaluation, run the script:

python metric_cal.py

Note: If the csv file recorded_metrics.csv exists before running the evaluation script, the metrics from two evaluations mix. To avoid this, you can delete or rename the recorded_metrics.csv manually everytime before evaluation.