13 ms·

I think you are putting too much trust in how good OpenAI is at bigger strategy right now. They have a very push focused strategy that fits with the limitations

by cc81 8y ago

I think you are putting too much trust in how good OpenAI is at bigger strategy right now. They have a very push focused strategy that fits with the limitations they have put on the game. When OpenAI wins against human opponents it seems mostly be because it is so much better at cooperating and can jump the enemy so quickly. That combined with continued ferrying of regen that would not work in a normal game.

shawn 8y ago
Mm, everyone said the same kind of thing last year. It's worth being skeptical, but I'd rather be a true believer and be proven wrong. It seems like betting against technology to say that they can't beat the top dota team in a fair fight within a couple years.
- Benjammer 8y ago
  I think the argument people used last year is still the same argument though. The 1v1 SF mid was all about computer-precision mechanical skill with hitting creeps and HP management. People said the AI would never be able to match pro-level meta/team strategies. While OpenAI Five is insanely impressive, it's still very glaringly obvious that their range of play styles is very restricted. Many of the pros, even ones who have played against (read: got crushed by) OpenAI Five, still think it's very far away for the bot to out draft a pro team with the full ranked game rules and hero pool. The push heavy death ball strategy is pretty much optimal for the high-precision, perfectly coordinated, mechanical fighting skill the bots have. They get a few kills using the mechanical skill and then they group up and press their advantage as hard as they can. It seems like all the rules/mechanics they are still working on are the more abstract out-of-the-box stuff that evolved to deal with these types of 5-man "teamfight" hero lineups... (warding, game-contextual item builds, courier management, full hero pool with all the "rat" split pushing heroes, etc.)
  - roenxi 8y ago
    > Many of the pros, even ones who have played against (read: got crushed by) OpenAI Five, still think it's very far away for the bot to out draft a pro team with the full ranked game rules and hero pool. Neither AI experts or DotA experts have the necessary background to predict how close anything is to any level of future performance. DotA players are probably the worst to ask, because the AI is already stronger than they are at a subset of the game and they have no insight into how quickly that subset could be generalised into other aspects. OpenAI has a list of something like 7 restrictions. Nobody has any real idea of how quickly these restrictions can be lifted once the OpenAI team has an AI that has mastered the game with those restrictions. Eg, 5 invulnerable couriers is obviously a huge restriction - but once OpenAI knows it gains a benefit by ordering items, how difficult is it to lift that restriction? Nobody knows. Might be easy, might be hard.
    - mellinoe 8y ago
      > OpenAI has a list of something like 7 restrictions. Some of those are very major, though, and I think the word "restriction" is a bit misleading. Having 5 invulnerable couriers is not really a "restriction" in that it limits or simplifies parts of the game -- it's just a fundamentally different mechanic that changes the way the game can be played. > DotA players are probably the worst to ask, because the AI is already stronger than they are at a subset of the game and they have no insight into how quickly that subset could be generalised into other aspects. I think that's a little unfair. Most folks have been pointing out that OpenAI's current momentum-based "deathball" strategy seems to fall apart without infinite regen and a limited hero pool, both facilitated by the current set of restrictions. I'd agree that nobody really knows how well OpenAI will adapt to the full game, but I disagree that the criticisms I've seen are meritless. OpenAI's current level of play is definitely impressive, but I think there's still room for skepticism given the current restrictions. I (and I think a lot of others) would be pretty disappointed if the TI showmatch happened with the turbo mode couriers still enabled.
      - roenxi 8y ago
        > I think that's a little unfair. Most folks have been pointing out that OpenAI's current momentum-based "deathball" strategy seems to fall apart without infinite regen and a limited hero pool, both facilitated by the current set of restrictions. People said similar things about Go AIs and ko fights. And in the end it turned out that neural networks handled kos fine but ladders were a challenge. On the deathball strategy in particular, consider that we expect a superhuman DotA AI to change the DotA metagame, so playing off-meta doesn't tell us anything. AlphaGo would invade 3-3 point a lot more enthusiastically than a human player. This was considered a classic beginner mistake for many years; now the theory has been readjusted to cope with the fact that AlphaGo stuck with it and just considered it a good move. We can safely say that the courier change has made a deathball strategy more powerful and it seems quite likely it is not an optimum strategy. But we can't be sure until OpenAI tests it, and we absolutely can't be sure that OpenAI won't just learn a new style when the conditions change. The criticisms have merit, but nobody has enough data predict anything about the future. Particularly a professional DotA player.

by cc81 8y ago

shawn 8y ago

Mm, everyone said the same kind of thing last year. It's worth being skeptical, but I'd rather be a true believer and be proven wrong. It seems like betting against technology to say that they can't beat the top dota team in a fair fight within a couple years.

Benjammer 8y ago
I think the argument people used last year is still the same argument though. The 1v1 SF mid was all about computer-precision mechanical skill with hitting creeps and HP management. People said the AI would never be able to match pro-level meta/team strategies. While OpenAI Five is insanely impressive, it's still very glaringly obvious that their range of play styles is very restricted. Many of the pros, even ones who have played against (read: got crushed by) OpenAI Five, still think it's very far away for the bot to out draft a pro team with the full ranked game rules and hero pool. The push heavy death ball strategy is pretty much optimal for the high-precision, perfectly coordinated, mechanical fighting skill the bots have. They get a few kills using the mechanical skill and then they group up and press their advantage as hard as they can. It seems like all the rules/mechanics they are still working on are the more abstract out-of-the-box stuff that evolved to deal with these types of 5-man "teamfight" hero lineups... (warding, game-contextual item builds, courier management, full hero pool with all the "rat" split pushing heroes, etc.)
- roenxi 8y ago
  > Many of the pros, even ones who have played against (read: got crushed by) OpenAI Five, still think it's very far away for the bot to out draft a pro team with the full ranked game rules and hero pool. Neither AI experts or DotA experts have the necessary background to predict how close anything is to any level of future performance. DotA players are probably the worst to ask, because the AI is already stronger than they are at a subset of the game and they have no insight into how quickly that subset could be generalised into other aspects. OpenAI has a list of something like 7 restrictions. Nobody has any real idea of how quickly these restrictions can be lifted once the OpenAI team has an AI that has mastered the game with those restrictions. Eg, 5 invulnerable couriers is obviously a huge restriction - but once OpenAI knows it gains a benefit by ordering items, how difficult is it to lift that restriction? Nobody knows. Might be easy, might be hard.
  - mellinoe 8y ago
    > OpenAI has a list of something like 7 restrictions. Some of those are very major, though, and I think the word "restriction" is a bit misleading. Having 5 invulnerable couriers is not really a "restriction" in that it limits or simplifies parts of the game -- it's just a fundamentally different mechanic that changes the way the game can be played. > DotA players are probably the worst to ask, because the AI is already stronger than they are at a subset of the game and they have no insight into how quickly that subset could be generalised into other aspects. I think that's a little unfair. Most folks have been pointing out that OpenAI's current momentum-based "deathball" strategy seems to fall apart without infinite regen and a limited hero pool, both facilitated by the current set of restrictions. I'd agree that nobody really knows how well OpenAI will adapt to the full game, but I disagree that the criticisms I've seen are meritless. OpenAI's current level of play is definitely impressive, but I think there's still room for skepticism given the current restrictions. I (and I think a lot of others) would be pretty disappointed if the TI showmatch happened with the turbo mode couriers still enabled.
    - roenxi 8y ago
      > I think that's a little unfair. Most folks have been pointing out that OpenAI's current momentum-based "deathball" strategy seems to fall apart without infinite regen and a limited hero pool, both facilitated by the current set of restrictions. People said similar things about Go AIs and ko fights. And in the end it turned out that neural networks handled kos fine but ladders were a challenge. On the deathball strategy in particular, consider that we expect a superhuman DotA AI to change the DotA metagame, so playing off-meta doesn't tell us anything. AlphaGo would invade 3-3 point a lot more enthusiastically than a human player. This was considered a classic beginner mistake for many years; now the theory has been readjusted to cope with the fact that AlphaGo stuck with it and just considered it a good move. We can safely say that the courier change has made a deathball strategy more powerful and it seems quite likely it is not an optimum strategy. But we can't be sure until OpenAI tests it, and we absolutely can't be sure that OpenAI won't just learn a new style when the conditions change. The criticisms have merit, but nobody has enough data predict anything about the future. Particularly a professional DotA player.