swstarlab / GACE-GDAN

Goal-Aware Cross-Entropy for Multi-Target Reinforcement Learning (NeurIPS'21) [arxiv]

This repository contains source code of our paper "Goal-Aware Cross-Entropy for Multi-Target Reinforcement Learning" at NeurIPS 2021.

This code contains our method, GACE&GDAN, for Visual Navigation tasks.

Dependencies

• python 3
• pytorch 1.7 +
• tensorboard 2.4
• numpy 1.17
• setproctitle 1.2
• Multi-target Visual Navigation environments (link)

Run

Before you run this script, please check params.py for allocating your GPU properly. Particularly, referring to below parameters,

self.gpu_ids_train = [0,1]

and

self.gpu_ids_test = [0,1]

indicate which GPUs to allocate for training and evaluating, respectively. If you have only one, set these parameters to [0]. Otherwise, you may allocate more GPUs.

Please make sure that self.num_training_process is set according to the number of CPU cores and the amount of GPU memory.

When you are ready, run the script to start the training:

python main.py

Citation

If you find our research helpful, please consider citing our paper,

@article{kim2021goal,
  title={Goal-Aware Cross-Entropy for Multi-Target Reinforcement Learning},
  author={Kim, Kibeom and Lee, Min Whoo and Kim, Yoonsung and Ryu, JeHwan and Lee, Minsu and Zhang, Byoung-Tak},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}