![]() Its popularity has arisen to a legendary status, spawning themed merchandise even in 2014. One of the most iconic video game systems for more than a single generation of gamers is the Nintendo Entertainment System. Review by Another World – Completed 7/30/14 Introduction ![]() If DM announced tomorrow that MuZero+VQ-VAE had defeated AS and a bunch of human pros, people would be extremely shocked, even as they scrambled to tweet about how they saw it coming all along & also are very offended by the waste of CO2 and how unreplicable it is on a grad student budget.Worldwide sales by:, Emere.es,, Neotienda.es,, .uk, Īlso Known As: Everdrive NES, N8, EN8, EverDrive N8, Everdrive-N8, EverDrive-N8, EverDrive NES (VQ-VAE and related generative modeling approaches are promising in this regard. Now, applying MuZero might be powerful to solve SC2 with much more compute, but that is still not something many DRL researchers would give you large odds on, and you still probably need to come up with some ways to abstract it and condense the game tree into something tractable for planning over, especially with the partial information making modeling the future much harder. Even with the very interesting AlphaStar League partially solving the self-play diversity and exploration problem, it's not at all obvious that just sinking in another 10-100x TPU-time into AlphaStar would patch up all of the errors and poor strategy AS exhibited - in fact, I'd bet heavily against it, since spending a large-but-still-feasible amount of compute doesn't appear to have fixed OA5 or AlphaGo's similar-looking problems, but AlphaZero was required. I'd say SC2 is still in the "inherently creative" part. At some point you'd know enough to be able to predict the next piece with 100% accuracy. I suppose one could extend this to be a 3-dimensional lookup table with the probabilities of the next pieces given the last two pieces, or to extend it to 4 or 5 or (if you had infinite resources) 100. ![]() There is some interesting code around the RNG, though: apparently the RNG does make certain piece sequences more likely than others, and there's a lookup table for the probability of the next piece given the current piece: Looking at the github repo, it looks like it's actually more of a classical AI doing traditional game tree search. I wondered if this was using a machine-learning style AI and it was, among other things, learning the state of the random number generator so it could predict pieces accurately more than one turn out? (And if not, how would you prove that it wasn't? Perhaps tweak the game RNG and see if the AI performs badly?) ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |