This project is used for the Generic Video Game Competition (GVGAI) Learning Competition since the year 2019. For more about the competition legs, rules and rank, please visite the AI in Games website, maintained by Hao Tong and Jialin Liu.

Disclamer

This project is forked from GVGAI Gym, which is an OpenAI Gym environment for games written in the Video Game Description Language (VGDL), including the GVGAI framework.

Please refer to the paper Deep Reinforcement Learning for General Video Game AI for more about the GVGAI GYM framework and some initial results of Reinforcement Learning (RL) agents. This paper should be cited if code from this project or the original GVGAI GYM project is used in any way:

@inproceedings{torrado2018deep,
  title={Deep Reinforcement Learning for General Video Game AI},
  author={Torrado, Ruben Rodriguez and Bontrager, Philip and Togelius, Julian and Liu, Jialin and Perez-Liebana, Diego},
  booktitle={Computational Intelligence and Games (CIG), 2018 IEEE Conference on},
  year={2018},
  organization={IEEE}
}

GVGAI Learning Competition in 2021

Two competition legs at IEEE CoG2021: http://aingames.cn/gvgai/cog2021/

Performance of some baseline planning agents giving a forward model

Some representive planning agents from the GVGAI Single-player Planning Competition are used as references in this test. Note that those agents have access to the game forward model, thus, the test was performed under planning setting instead of learning setting.
Theoretical maximum refer to the maximum score that one can get if playing optimally.

Game-level\ Planing agent	RHEA	MCTS	OLETS	Random	Theoretical maximum
greedymouse-lv0	21.4(0)	29(0)	57.8(4)	-4.7(0)	98
greedymouse-lv1	12.7(0)	20(0)	26.1(0)	-2.23(0)	67
bravekeeper-lv0	38.7(24)	40.8(30)	47(28)	8.5(0)	100
bravekeeper-lv1	36.2(12)	51.7(26)	50.8(24)	5.0(0)	90
trappedhero-lv0	1.7(0)	2.0(1)	13.3(25)	1.5(1)	15
trappedhero-lv1	0(0)	0.17(0)	10.2(17)	0(0)	15

Table 1: Average score (wins) of different planning agents on each game level over 30 independent trials.

Latest Updates

Installation

Way 1: using Docker

Please refer to the step-by-step guidelines for setting up the framework and RL baselines with Docker (using GPU or CPU only).

Way 2: git clone

Clone this repository to your local machine
Run pip install -e <package-location> to install the package
Install a Java compiler(e.g. sudo apt install openjdk-9-jdk-headless)

Requirements

Anaconda:　The version published after 2019.10 is recomended
Java:　JDK 9 is recommended
Python:　The version Python3 (3.6 or 3.7 are recomended) is acceptable. (Python2 can't be used!!!)