4 ms·
How to train your decision-making AIs
- optimalsolver 5y agoWas DeepMind's new Player of Games AI tested on Montezuma's Revenge? That seems to be the Atari game that gives ML agents the most trouble.
- andreyk 5y agoI don't believe so, as it's more about competitive games and about incomplete information. They did release Agent 57 last year, and that got better than human results on all Atari games (though still not sample efficiently). More recently there was EfficientZero, an extension to MuZero, which is pretty cool as it is both sample efficient and quite performant, though I don't think it can do Montezuma. https://arxiv.org/abs/2111.00210 https://arxiv.org/abs/2111.00210
- gwern 5y agoExploration is a different sort of problem, so neither PoG nor EfficientZero immediately gives you a solution for that. (They probably can do it well because their model-based nature gives them ways to quantify uncertainty for exploration. But they haven't yet.)
- ece 5y agoRelated: https://www.theverge.com/2016/6/9/11893002/google-ai-deepmind-atari-montezumas-revenge https://www.theverge.com/2016/6/9/11893002/google-ai-deepmin...
- andreyk 5y agoHere's a link to the original paper for those who might prefer it: https://link.springer.com/article/10.1007/s10458-021-09514-w https://link.springer.com/article/10.1007/s10458-021-09514-w This is a slightly edited and updated version of it.