Code for paper Perceiving the World: Question-guided Reinforcement Learning for Text-based Games
Yunqiu Xu, Meng Fang, Ling Chen, Yali Du, Joey Tianyi Zhou and Chengqi Zhang
- An overview of the decision making process:
- Model architecture:
-
Our code depends heavily on xingdi-eric-yuan/GATA-public. The additional dependencies could be found at requirements.txt
-
Download the word embeddings:
wget "https://bit.ly/2U3Mde2"
- Datasets for pre-training the task selector and the action validator are provided at this link, other datasets could be downloaded at:
# AP
wget https://aka.ms/twkg/ap.0.2.zip
# RL
wget https://aka.ms/twkg/rl.0.2.zip
-
Modify the paths within the config files, e.g. "word_embedding_path"
-
Action prediction (providing initialization for the encoders):
python train_ap.py config/config_pretrainAP.yaml
- Task selector (pre-training phase):
python train_vt.py config/config_pretrainVT.yaml
- Action validator (pre-training phase):
python train_va.py config/config_pretrainVA.yaml
- Action selector (reinforcement learning phase):
# Medium games
python train_rl_medium.py config/config_trainRL_medium.yaml
# Hard games
python train_rl_hard.py config/config_trainRL_hard.yaml
@inproceedings{xu-etal-2022-perceiving,
title = "Perceiving the World: Question-guided Reinforcement Learning for Text-based Games",
author = "Xu, Yunqiu and
Fang, Meng and
Chen, Ling and
Du, Yali and
Zhou, Joey and
Zhang, Chengqi",
booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = may,
year = "2022",
address = "Dublin, Ireland",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.acl-long.41",
doi = "10.18653/v1/2022.acl-long.41",
pages = "538--560"
}