Skip to content
master
Switch branches/tags
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
Nov 6, 2018
Nov 6, 2018
Jan 3, 2018
Jan 29, 2020
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018
Nov 6, 2018

This is a source code for AAAI 2019 paper Classification with Costly Features using Deep Reinforcement Learning wrote by Jaromír Janisch, Tomáš Pevný and Viliam Lisý: paper / slides / poster / code / blog

Updated version available: There is an enhanced version of the article under name Classification with Costly Features as a Sequential Decision-Making Problem (paper), which analyzes more settings (hard budget, lagrangian optimization of lambda and missing features). The code is available in the lagrange branch of this repository.

Cite as:

@inproceedings{janisch2019classification,
  title={Classification with Costly Features using Deep Reinforcement Learning},
  author={Janisch, Jaromír and Pevný, Tomáš and Lisý, Viliam},
  booktitle={AAAI Conference on Artificial Intelligence},
  year={2019}
}

Prerequisites:

  • cuda capable hardware
  • ubuntu 16.04
  • cuda 8/9
  • python 3.6 (numpy, pandas, pytorch 0.4)

Usage:

  • use tools tools/conv_*.py to prepare datasets; read the headers of those files; data is expected to be in ../data
  • pretrained HPC models are in trained_hpc, or you can use tools/hpc_svm.py to recreate them; they are needed in ../data
  • run python3.6 main.py --dataset [dataset] --flambda [lambda] --use_hpc [0|1] --pretrain [0|1], choose dataset from config_datasets/
  • the run will create multiple log files run*.dat
  • you can use octave or matlab to analyze them with tools/debug.m
  • you can also evaluate the agent on the test set with eval.py --dataset [dataset] --flambda [lambda]

About

Source code for paper Classification with Costly Features using Deep Reinforcement Learning.

Topics

Resources

License

Releases

No releases published

Packages

No packages published