Comment by abecedarius | Hacker News Reader

Comment by abecedarius | Hacker News Reader

Parent Full thread

The same company, DeepMind, already published about reinforcement learners reaching superhuman performance on many realtime games -- old Atari ones.

I know there's a lot of hype and ignorant confident opinions being expressed, but this sort of response seems really strange to me: in the last decade we went from nobody having a clue how we might automate Go in this lifetime, to 2014 when an article on prospects for Go by a premier researcher on game AI (https://news.ycombinator.com/item?id=11290112) did not even mention neural nets, to professional play last fall, to 5 months later crushing a top player and making high-level innovations, in a game said to be among the most deep and beautiful ever invented... and the most salient points are about how trivial it all is?

The same company, DeepMind, already published about reinforcement learners reaching superhuman performance on many realtime games -- old Atari ones

This is not remotely similar to league of legends or starcraft.

I would think LoL and Starcraft would be easier in terms of difficulty than something like Go, because they both have the fast twitch aspect of play. An automated player would have the advantage of perfect situational awareness of the mini map and consistency in character placement and so forth. This is a big advantsge in games where a single misclick or a missed visual can cost a match.

Strategy is what matters here. That's what I was getting at with the last part of my first message. We can make bots that do certain aspects of the game flawlessly and we will need to find a way to integrate them.

It's also very understandable from my perspective that we would have been able to surpass the best players given that monte carlo bots are already around 5-6 dan skill level.

Don't forget that monte carlo bots were around 5-6 _amateur_ dan skill level. It is _very_ different from 9 dan pro.

The difference from 5 amateur dan to 9 dan pro is much smaller than the difference between 10 kyu and 5 amateur dan. I figured google with billions in resources thrown at monte carlo alone would be enough to contend with pros. We didn't exactly have huge organizations creating these bots before.

Reply on news.ycombinator.com