Skip to content

Latest commit

 

History

History
37 lines (26 loc) · 2.1 KB

README.md

File metadata and controls

37 lines (26 loc) · 2.1 KB
Raw

XTX: eXploit - Then - eXplore

Project page: https://sites.google.com/princeton.edu/xtx

Requirements

First clone this repo using git clone https://github.com/princeton-nlp/XTX.git

Please create two conda environments as follows:

  1. conda env create -f yml_envs/jericho-wt.yml
    a. conda activate jericho-wt
    b. pip install git+https://github.com/jens321/jericho.git@iclr
  2. conda env create -f yml_envs/jericho-no-wt.yml

The first set of commands will create a conda environment called jericho-wt which has added actions to the game grammar for specific games (see games with * in the paper). The second command will create another conda environment called jericho-no-wt which installs an unmodified version of the Jericho library.

Training

All code can be run from the root folder of this project. Please follow the commands below for each specific model:

  • XTX: sh scripts/run_xtx.sh
  • XTX (no-mix): sh scripts/run_xtx_no_mix.sh
  • XTX (uniform): sh scrtips/run_xtx_uniform.sh
  • XTX ($\lambda$ = 0, 0.5, or 1): sh scripts/run_xtx_ablation.sh
  • INV DY: sh scripts/run_inv_dy.sh
  • DRRN: sh scripts/run_drrn.sh

Notes

  • You can use analysis/sample_env.py for quickly playing around with a sample Jericho environment. Run it using python3 -m analysis.sample_env.

  • You can use analysis/augment_wt.py for generating the missing action candidates that can be added to the game grammar (games with * in the paper). Run it using python3 -m analysis.augment_wt.

  • Note that all models should finish within a day or two given 1 gpu and 8 cpus, except for games where Jericho's valid action handicap is slow (e.g. Library, Dragon). Since Jericho's valid action handicap heavily relies on parallelization, increasing the number of cpus also results in good speedups (e.g. 8 -> 16).

Acknowledgements

We used Weights & Biases for experiment tracking and visualizations to develop insights for this paper.

Some of the code borrows from the TDQN repo.

For any questions please contact Jens Tuyls (jtuyls@princeton.edu).