@jackclarkSF @Miles_Brundage What does @OpenAI’s Dota 2 victory say about the long-running debates about ML vs symbolic reasoning techniques? Do the OpenAI5 use a hybridized approach or is this a clear win for deep RL?
2
1
See - nitter.vloup.ch/polynoamial/stat… perhaps Noam or others will have more to add :)
Welp, I'm surprised. Congrats to @OpenAI on their win over top Dota2 team @OGesports!
2
1
My hot take: OpenAI Five uses model-free RL so it's arguably a win for that end of things vs. the more real-time search/planning + explicit modeling of the world end of things, though the fact that it took 45k years to learn that performance shows there is a long way to go in AI!
2
2
As an outsider, I don't have any stake in these arguments. And perhaps its my ignorance at work here, but I'm far more impressed by what deep RL can do than by what it cannot do. Obviously lots of limitations, but deep RL seems like a very powerful tool.
2
1
How does this achievement -- and its limitations -- intersect with the "bitter/better lesson of compute" argument and the value of generalizable methods + compute? incompleteideas.net/IncIdeas…
1
One last question: How generalizable is the method behind #OpenAI5 to other RTS games, or is the network highly tailored to Dota 2?
1
Highly tailored, expect no meaningful generalization, look at their network design, its entirely dota-api-specific: d4mucfpksywv.cloudfront.net/…
1
1
The learning system is fully general-purpose — that's what powers this: blog.openai.com/learning-dex….
Network is basically a bunch of Dota-specific (but straightforward) embeddings, etc to get the game state to run through an LSTM.
The training Data is totally Dota-specific.
Apr 16, 2019 · 6:54 PM UTC
3