Sign In with Google
Sign In with Facebook
GridWorlds in the browser (alpha)
AI safety gridworlds
Absent supervisor
Boat race
Distributional shift
Friends and foes
Island navigation
Safe interruptibility
Side effects sokoban
Tomato watering
Whisky gold
Gridworlds
Aperture
Apprehend
Extraterrestrial marauders
Fluvial natation
Hello world
Scrolly maze
Shockwave
Warehouse manager
Chain walk
Cliff walk
Four rooms
Your past best scores:
Best reward:
Show leaderboard ...
×
Leaderboard
Agent code
∎
❚❚
►
Environment code
(read-only)
The goal
Project for popularisation of AI safety topics through competitions and gamification
Author
Roland Pihlakas
Special thanks to
Eero Ränik
Major components
AI Safety Gridworlds by DeepMind
XTerm.JS by SourceLair
Docker
Monaco editor by Microsoft
CloudPickle by CloudPipe
Isso by Martin Zimmermann
Miniconda by Continuum Analytics
Python 3.5
Python 2.7
Node.JS
MongoDB
CentOS
Roadmap:
- Support for uploading and downloading the agent data file.
- Performance and stability improvements.
- General discussion forum.
- Ability to save human generated moves to training data files.
- Environment editor (a la
The Incredible Machine
).
- Additional environments.
- Sample agents.
- Embedding support.
- Step by step lessons.
- Help texts.
-
Environment for illustration of fundamental limits of attention.
-
Environment for illustration of whitelisting approach.
-
Environment for illustration of homeostatic diminishing returns and conjunctive goals approach.
- GitHub integration.
- Support for programming the agents in
Scratch visual programming language
.