When we lost at The International (100% vs pro teams), they said it was because Five can’t do strategy. So we trained for longer. When we lose (0.7% vs the entire Internet), they say it’s because Five can’t adapt. If we train for longer...??? nitter.vloup.ch/metaversefight/s…

Apr 21, 2019 · 1:59 PM UTC

8
6
83
One counterpoint: our scripted components are a clear weakness that won’t change with training. But that’s just a handicap. There are still *many* aspects of the game, like invisibility, which are learnable. Our experience tells us never to discount the power of more training.
4
3
56
Replying to @gdb
I'd imagine the biggest challenge is making it adaptable to changes from patches (game/hero/item adjustments rather than additions of heroes). Maybe if you increased the params and trained it throughout multiple patches, but obvs very expensive. Any plans/strategies for that?
1
Five already does adjust automatically between patches, as described here: openai.com/blog/how-to-train…
Replying to @gdb
Shouldn’t there be fundamental limit to depth of behavioral complexity for a given size and configuration of neural network? And human teams still winning could indicate either of oversized network with overfit or undersized network with insufficient capacity. How to distinguish?
1
1
Replying to @gdb
The question is: does training time grows exponentially though? As we know there are 5 times more heroes and there is a big pool of heroes with high complexity (invoker, storm, meepo, etc) also there are mechanisms like illusions. Don't get me wrong I am very excited about
1
1
Replying to @gdb
TI12 champs