natnij/draco

source code for INFOCOM paper

Multi-Agent Distributed Reinforcement Learning for Making Decentralized Offloading Decisions

Background

This is the source code for a paper submission.

Requirements

pandas==0.24.2

matplotlib==3.1.2

numpy==1.16.4

scipy==1.2.1

seaborn==0.9.0

torch==1.4.0

Mesa==0.8.6

Install the package

Go to the virtual environment and the root v2x folder, run:

foo@bar:~/v2x$ python -m pip install -r requirements.txt
foo@bar:~/v2x$ python -m pip install -e .

Generate the performance charts in paper

extract all files in logs folder (use "extract here"), create a new graphs folder, and run:

foo@bar:~/v2x$ python output.py

The charts will be created in the graphs folder.

Run the code in predefined modes

Create a new models folder, the models will be saved here.

1. To run the DRACO algorithm in training mode:

   foo@bar:~/v2x$ python run.py draco

2. To permit rebidding,add keyword "rebid":

   foo@bar:~/v2x$ python run.py draco_rebid

   Maximum permitted rebidding is specified in the config file in v2x/config folder, parameter nrRebid.

3. To run in evaluation mode, add keyword "eval":

   foo@bar:~/v2x$ python run.py draco_rebid_eval

   Performance output will be automatically created in the logs folder. If there are existing files in the folder with the same name, the new results will be appended to the end.

4. To run benchmark RIAL:

   foo@bar:~/v2x$ python run.py rial

5. To permit rebidding for RIAL, add keyword "rebid":

   foo@bar:~/v2x$ python run.py rial_rebid