The Robot Soccer Goal environment [Masson et al. 2016] uses a parameterised action space and continuous state space. The task involves an agent learning to kick a ball past a keeper. Three actions are available to the agent:
- kick-to(x,y)
- shoot-goal-left(y)
- shoot-goal-right(y)
A reward of 50 is given for a successful goal, and -distance(ball, goal)
otherwise. An episode terminate if the ball enters the goals, is captured by the keeper, or leaves the play area.
This code is a port of https://github.com/WarwickMasson/aaai-goal to use the OpenAI Gym framework.
- Python 3.5+ (tested with 3.5 and 3.6)
- gym 0.10.5
- pygame 1.9.4
- numpy
Install this as any other OpenAI gym environment:
git clone https://github.com/cycraig/gym-goal
cd gym-goal
pip install -e '.[gym-goal]'
or
pip install -e git+https://github.com/cycraig/gym-goal#egg=gym_goal
import gym
import gym_goal
env = gym.make('Goal-v0')
See https://github.com/cycraig/MP-DQN for an example on how to make an agent for this environment.
If you use this domain in your research, please cite the original author:
@inproceedings{Masson2016ParamActions,
author = {Masson, Warwick and Ranchod, Pravesh and Konidaris, George},
title = {Reinforcement Learning with Parameterized Actions},
booktitle = {Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence},
year = {2016},
location = {Phoenix, Arizona},
pages = {1934--1940},
numpages = {7},
publisher = {AAAI Press},
}
You may also consider citing the following paper:
@article{bester2019mpdqn,
author = {Bester, Craig J. and James, Steven D. and Konidaris, George D.},
title = {Multi-Pass {Q}-Networks for Deep Reinforcement Learning with Parameterised Action Spaces},
journal = {arXiv preprint arXiv:1905.04388},
year = {2019},
archivePrefix = {arXiv},
eprinttype = {arxiv},
eprint = {1905.04388},
url = {http://arxiv.org/abs/1905.04388},
}