Imbalanced Classification with Deep Reinforcement Learning.
This repository contains an (Double) Deep Q-Network implementation of binary classification on unbalanced datasets using TensorFlow 2.3+ and TF Agents 0.6+. The Double DQN as published in this paper by van Hasselt et al. (2015) is using a custom environment based on this paper by Lin, Chen & Qi (2019).
Example scripts on the Mnist, Fashion Mnist, Credit Card Fraud and Titanic datasets can be found in the ./imbDRL/examples/ddqn/ folder.
The following results are collected with the scripts in the appendix: imbDRLAppendix. Experiments conducted on the latest release of imbDRL and based on this paper by Lin, Chen & Qi (2019).
- Python 3.7+
- The required packages as listed in:
requirements.txt - Logs are by default saved in
./logs/ - Trained models are by default saved in
./models/ - Optional:
./data/folder located at the root of this repository.- This folder must contain
creditcard.csvdownloaded from Kaggle if you would like to use the Credit Card Fraud dataset. - Note:
creditcard.csvneeds to be split in a seperate train and test file. Please use the functionimbDRL.utils.split_csv
- This folder must contain
Install via pip:
pip install imbDRL
Run any of the following scripts:
python .\imbDRL\examples\ddqn\train_credit.pypython .\imbDRL\examples\ddqn\train_famnist.pypython .\imbDRL\examples\ddqn\train_mnist.pypython .\imbDRL\examples\ddqn\train_titanic.py
To enable TensorBoard, run tensorboard --logdir logs
Extra arguments are handled with the ./tox.ini file.
- Pytest:
python -m pytest - Flake8:
flake8 - Coverage can be found in the generated
./htmlcovfolder
The appendix can be found in the imbDRLAppendix repository.