Awesome

Monte-Carlo Planning and Learning with Language Action Value Estimates

This repository is the implementation of "Monte-Carlo Planning and Learning with Language Action Value Estimates"

Requirements

To install requirements:

conda env create -f environment.yml
conda activate mc-lave-rl
python -m spacy download en_core_web_lg
python -m spacy download en

Monte-Carlo Tree Search (MCTS)

To run the data collection with MCTS for initial iteration, run this command with :

python run_planning.py --seed=$SEED --trial=0

(For each planning, use --seed option with different index)

All experiments in the paper were conducted with Google Cloud Platform(GCP) and HTCondor to run in parallel. If you want to run the experiment in parallel, we reccomend to set up the HTCondor and environment on your GCP account.

Merge planning trajectories and experience replays

To merge the data for learning the Q-network and policy, run this command:

python merge_plans.py

Train the Q-network and policy

To run the Q-network and policy with collected data, run this command:

(train Q-network) python run_learning.py --mode=0
(train policy) python run_learning.py --mode 1

Monte-Carlo Planning with Language Action Value Estimates (MC-LAVE)

To run the MC-LAVE planning with trained Q-network and policy, run this command with :

python run_planning.py --seed=$SEED --trial=1

(For each iteration, use --trial option with iteration number)

Citation

If this repository helps you in your academic research, you are encouraged to cite our paper. Here is an example bibtex:

@inproceedings{jang2021montecarlo,
  title={Monte-Carlo Planning and Learning with Language Action Value Estimates},
  author={Youngsoo Jang and Seokin Seo and Jongmin Lee and Kee-Eung Kim},
  booktitle={International Conference on Learning Representations},
  year={2021},
  url={https://openreview.net/forum?id=7_G8JySGecm}
}

Acknowledgement

This code is adapted and modified upon the code github of AAAI 2020 paper "Interactive Fiction Games: A Colossal Adventure". We appreciate their released dataset and code which are very helpful to our research.