OpenAI Five can't learn from humans, but humans can learn from Five. Cool analysis:
We looked at some of the things @OpenAI did to out maneuver Team Human at the benchmarking event last weekend.
youtube.com/rTtg_DdOqPQ
2
12
1
70
Glad you enjoyed it Greg! We are really excited to see how the project evolves over the next year! The drafting in particular was very intriguing. Were the rankings/percentages for each hero just data that OpenAI Five developed itself through self play?
1
5
(Note that during training it just plays randomized lineups, so it’s not likely it’s played many games with all the lineups — instead it must generalize from what it has experienced.)
Aug 9, 2018 · 4:08 PM UTC
4

