4 ms·

While this is an interesting breakthrough for AI, let's remember that Go is only limitless in the exhaustive possibilities for every response move in sequence f

While this is an interesting breakthrough for AI, let's remember that Go is only limitless in the exhaustive possibilities for every response move in sequence for the game. So many moves in response to a move are just not good and invalid you can do extensive pruning of search trees rather easily. That's why monte carlo algorithms work so well as it is. They don't beat the best of the best but I mean they still beat like literally 90% of Go Players. But the game's rules are simple, the total storage of state needed for a game is rather small, so in terms of utilizing training data it's quite easy. You can fit an entire Go game in like 2 kilobytes of memory. There's only two types of stones, and they have no meaning other than being different than each other. We would need more breakthroughs for real-time games. Compared to Go, they have easily gigabytes of potentially meaningful data PER game. Not only is storing millions and millions of games becoming an issue, processing them is becoming an issue as well. And in many of those games when strategy changes depending on which character or map you're on, it quickly balloons out in a manner that is just not reasonable. Maybe the next step will be specialized AIs that tackle small more easily calculable components of real-time games and then combining those together to make something decent.

abecedarius 11y ago
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 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?
- Madmallard 11y ago
  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.
  - stone-monkey 11y ago
    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.
    - Madmallard 11y ago
      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.
  - Umn44 11y ago
    I can see league of legends easily being AI'd. Starcraft is a lot more difficult because of the realtime nature of the game. League of legends was made for people with poor reflexes. I'd love to see an AI go at quake CTF game personally, or something more complex like UT2004 game modes.
- Madmallard 11y ago
  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.
  - shmel 11y ago
    Don't forget that monte carlo bots were around 5-6 _amateur_ dan skill level. It is _very_ different from 9 dan pro.
    - Madmallard 11y ago
      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.

abecedarius 11y ago

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 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?

Madmallard 11y ago
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.
- stone-monkey 11y ago
  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.
  - Madmallard 11y ago
    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.
- Umn44 11y ago
  I can see league of legends easily being AI'd. Starcraft is a lot more difficult because of the realtime nature of the game. League of legends was made for people with poor reflexes. I'd love to see an AI go at quake CTF game personally, or something more complex like UT2004 game modes.
Madmallard 11y ago
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.
- shmel 11y ago
  Don't forget that monte carlo bots were around 5-6 _amateur_ dan skill level. It is _very_ different from 9 dan pro.
  - Madmallard 11y ago
    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.