DeepNap

@Author Jingchu Liu, @Date Feb 17 2017

DeepNap is a deep reinforcement learning based sleeping control agent for base stations in mobile networks. This repository is maintained for review purposes of our paper: DeepNap: Data-Driven Base Station Sleeping Operations through Deep Reinforcement Learning

Requirements and Dependencies

All code are tested for Ubuntu 14.04.4 LTS with CUDA 7.5.17. The required Python packages are listed in requirements.txt. ipython Notebook is also required for visualization.

Setup

Clone the code and data from Github:

git clone https://github.com/zaxliu/deepnap.git

Installing Python dependencies (recommend using virtual environment):

cd project_home/
pip install -r requirements

Download and unzip data (need Internet connectivity to download data ~300MB)

cd project_home/
python setup.py

Try running the mimi-experiment to test if installation is sucessful:

cd project_home/experiments/
python run_mini.py

You should start seeing logging outputs in shell.

Reproducing Experimental Results

We provide code and data to reproduce the experimental results (figures and tables) in our paper.

Note you don't have to re-run all the experiments to see results. The setup routine will automatically download necessary datasets for visualization. You can find and reproduce all figures and tables used in our paper with this ipython notebook.

If you do want to repeat all experiments, we also provide a single script to re-run all the experiments that we tested. Note as this process will produce a large amount of log and intermediate data files, we recommend >500GB disk space researved for testing. Disk requirements can be relaxed to 20GB if you can manually delete all .log and index_*.log.csv files after you see the corresponding .reward file.

Also for speed considerations, we recommend using multi-core CPU, GPU, and >8G memory to fully leverage parallel execution. And be warned experiments are time-consuming - each experiment may take hours to finish. So be patient and have a drink, maybe.