A project to create randomized imperfect information extensive form games
Generalist AI Agents trained over a variety of games can outperform specialist AI Agents
(Deepmind SIMA).
This is an experimental idea to create randomized imperfect information extensive form games for the purpose of training a generalist AI Agent.
Scoping out general direction...
- Can one create/train the model to learn a multi-game action embedding, so actions across games can implicitly be compared?