These are chat archives for FreeCodeCamp/DataScience

Aug 2018
Alice Jiang
Aug 25 2018 00:46
Yeah, they've been making progress really fast. I heard the bots are playing ~400k games every day. I don't know how accurate that is but they definitely have a level of endurance that humans just don't lol.
I've been super obsessed with it because it's not a single system controlling the bots, it's five individual, isolated networks that have learned to work together, and their coordination is incredible. Not on level with the top teams in the world, but they beat out players from the 95th percentile easily, and dragged out the two games for quite a while, even after only having just a handful of days to add completely new aspects of the game that they previously hadn't been playing with.
I mostly follow professional league of legends and overwatch, so I don't know a lot about dota2, but if I understand the analysts, they were saying the biggest fault of the bots if the reaction time (300 ms I think) and troubles predicting how teams work
on an individual level they're really good. Most professional dota2 players have trained with OpenAI's 1v1 bots because the bots got better than humans, but it's not had as much time to figure out how teams work lol
Eric Leung
Aug 25 2018 03:10
@becausealice2 "five individual, isolated networks"?! That's creative. Interesting take on making machine work in some way with each other. There were those adversarial neural networks where they compete with each other, but to turn it around and make them cooperative (at least that's how I'm guessing how these networks are working) is quite fascinating. Mmm maybe I'll take a look at it a bit more.
Alice Jiang
Aug 25 2018 16:01
Yeah, they have as much insight into what each other is planning to do as any human teammates. One of the OpenAI engineers was saying they managed it in large part with a hyperparameter for selfishness. Their Youtube channel and blog are fascinating, for this and the robotics thing they're doing