A maze exploration game for AI agents.
- New game modes
- Terminal state for human, with switch for god-mode
- Zoomed scrolling view for human players
git clone https://github.com/mryellow/maze_explorer.git
cd maze_explorer
pip install -e .
pip install mazeexp
python standalone.py
--mode X # Mode number
--random # Execute random actions step-by-step via `act`
--step # Call the engine step-by-step via `step`
Apples and poison.
Based on Andrej Karpathy's Javascript environment
[
[min(wall_range, apple_range, poison_range), apple_range, poison_range],
...
]
avg(proximity) or 1
Agents don't like seeing walls, especially up close*1.1
Forward action bonus-10
collision with wall+5
collision with apple-6
collision with poison
- None
Explore the maze and make it back to spawn before battery runs out.
[wall_range, ..., battery]
-100
collision with wall-100
battery out
+1
exploration reward on first visit to tile and for each of it's open neighbours
The visited
state is not observable in environment and reward is generated by ground-truth.
Thus agent must keep it's own internal state and/or develop a policy which overcomes this unknown.
+200
reward goal state on reaching spawn tile- No futher exploration rewards
Spawn tile is no different to any other from agents perspective, must remember how to return to it or develop a policy which increases the likelihood of such.
- Wall collision
- Battery out
- Return to home goal
If you use Maze Explorer in your academic research, we would be grateful if you could cite it as follows:
@misc{mryellow2017mazeexplorer,
author = {Mr-Yellow},
title = {Maze Explorer: A maze exploration game for AI agents},
howpublished={Web page},
url = {https://github.com/mryellow/maze_explorer},
year = {2017}
}