Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Polytopia Simulation for AI Agents
We present a Polytopia simulation environment, explain the challenges of its 20,000‑action turn space, and discuss methods to train effective multi‑agent AI.
A simulation environment for a game called Polytopia to train AI agents on it and potentially utilize this system for AI Safety/Alignment research purposes. Google DeepMind actually once tried working on this exact project but they gave up because the action space in Polytopia was too large (we calculated about 20000 potential steps per agent per each turn, there are many agents playing at once, and each agent has to make a choice based on what they think the other agents will do). This is basically the problem we tried to deal with: How do we train good AI agents for tasks where the number of potential next actions is just too high?
Py4J bridges Java game engine to Python RL agents using Gym APIs.
Compose Email
Loading recent emails...