Page 2 of 4 FirstFirst 1 2 3 4 LastLast
Results 11 to 20 of 38

Thread: open-ai's 1v1 mid bot code open source?

  1. #11
    Quote Originally Posted by TheP1anoDentist View Post
    if it doesnt get open sourced that would be disappointing though. Im pretty sure if you're gonna call you're company Open<something>....it should prob be open source

    so do you think it's using image processing rather than in-game variables?

    also Im ok with what they did, I think its pretty cool.

    but when you have elon musk making statements like "OpenAI first ever to defeat world's best players in competitive eSports. Vastly more complex than traditional board games like chess & Go.", it spoils it.

    I dont get why he needs to pretend its bigger than it is, maybe it was just communication failure and he doesn't really understand the limits of the 1v1 bot.

    Considering it took 2 weeks to train that well. you've got 25x more time for 5v5. 25x I consider not a lot when a) the games last 5x as long. b) you're going to have orders of magnitude more input features.
    They are definately using in game variables, but my guess is that it is a custom API like this one.

    I think the tweet is not a mistake, but rather a deliberate misrepresentation of what was actually accomplished to make the project look good to the investors. Even ignoring the fact that the tweet and the article only came out seconds after the event, just listening to how those people were talking about this project during the event is enough for someone who knows advanced machine learning to understand that this is just a show to over exaggerate a relatively small accomplishment.

    And no, even assuming that all their claims are true, that does not mean that they are even remotely close to writing a 5v5 bot. Even for a bot that can actually play 1v1 (any hero vs any hero not just sf vs sf) they need around 10,000 times more time to train (even if it takes 1% of this, it takes around 4 years to train it), and 5v5 is exponentially harder than this. So basically with their current methods, I think it is impossible to do something like that!

    Edit: I don't think it will be open source, even though I like to be proven wrong and see that they have actually accoumplished what they are claiming (even though that is still not significant enough given their budget and resources)!
    Last edited by Platinum_dota2; 08-11-2017 at 09:43 PM.

  2. #12
    Basic Member BEST_'s Avatar
    Join Date
    Sep 2016
    Posts
    950
    How add this bot in dota 2? i want rape him. lul

  3. #13
    Basic Member
    Join Date
    Dec 2016
    Posts
    180
    Next Year 5vs5 ... Thats pretty cool if developers have a chance to raise their team, facing each other

    Hope valve developers share source code to us. If not, teaching or sharing knowledge inlua developments related to DOTA 2 will be good, at least

  4. #14
    Basic Member
    Join Date
    Jan 2014
    Posts
    57
    There are two points worth mentioning in these matches:

    First, the AI had access to play style and pattern of these players but the players didn't know anything about the pattern of the AI (Exactly like Deep-blue vs Kasporov matches in chess in early 2000 where Kasparov complained about not knowing the play style of Deep-blue).

    I bet if Dendi or Sumail allowed to play for a day against this AI, they could have beaten it easily.

    Second, the probability of accessing of the AI to "Unintended data". For example, currently the default AI monitors the key stroke of the player and the instant the player tapdown to cast a spell, they counter it. Its a kind of cheating (that is OK most of the time because AI is weak) but in this kind of scenarios, IF the AI has access to key stroke of the players, it could easily counter them.
    Last edited by fandogh; 08-12-2017 at 05:02 AM.

  5. #15
    @Platinum can you explain why creep aggro is currently impossible in bot api?

  6. #16
    Quote Originally Posted by TheP1anoDentist View Post
    @Platinum can you explain why creep aggro is currently impossible in bot api?
    It is not "impossible", but it can't be done "properly" since GetAttackTarget only works when the animation begins. So knowing what is the target of creeps before they start their attack animation is not a trivial task, but hey, their bot learned it should block creeps by itself from completely random moves, right? so it also should've learned how to agro creeps by analyzing their movement patterns! I really like to believe them, but it is hard to believe people when they write "I came up with an algorithm that learns how to stack a bunch of cubes" papers and suddenly claim that their algorithm can learn how to do brain surgery but they don't make their code public!

  7. #17
    Basic Member aveyo's Avatar
    Join Date
    Aug 2012
    Location
    EU West
    Posts
    2,927
    Quote Originally Posted by Platinum_dota2 View Post
    It is not "impossible", but it can't be done "properly" since GetAttackTarget only works when the animation begins. So knowing what is the target of creeps before they start their attack animation is not a trivial task, but hey, their bot learned it should block creeps by itself from completely random moves, right? so it also should've learned how to agro creeps by analyzing their movement patterns! I really like to believe them, but it is hard to believe people when they write "I came up with an algorithm that learns how to stack a bunch of cubes" papers and suddenly claim that their algorithm can learn how to do brain surgery but they don't make their code public!
    Soviet cheaters would like a word with you. Think Ensage. Think d2js. Think about your bots running that shit on top.
    Ofc they don't use the lame bot interface by itself.
    It's actually a trivial task to bridge vscript lua panorama js and dotakeys cfg together without any injected external tools - and then, azure is the limit
    But sure, I'm a little bullshit-resilient towards the early training open ai had - it looks like it had some assistance. Still impressive nonetheless.

  8. #18
    Basic Member
    Join Date
    Oct 2012
    Posts
    33
    I would like to see the source code as well, Im fairly confident that they used Universe which is OpenAI's neural network toolkit. It would not be surprising if they create a 5v5 full team of bots by next year considering the achievements of google's AI alphaGo which beat professional Go players and then OpenAIs poker bot which destroyed professional poker players.

    It was extremely disappointing that this bot only played a hero like shadowfiend and the player only could choose to play shadowfiend. The shadowfiend vs shadowfiend matchup basically ends after the first 30 secs because of the soul mechanic and as soon as a person dies; the lead against them becomes so overwhelming. I have worked with Tensorflow and Universe and built some cool AIs for different tasks, games and programs. I'm sure a dota bot wouldn't be hard to actually make with a team of 10 experienced developers but putting so many constraints of the players just seems like it defeats the purpose of making a bot like this.

    The bot should have knowledge of all of the heroes and items. It should be able to do more then just last hit, deny, cast raze and kill. For instance, if during a 1v1 match the player chooses to go bot instead of mid, the bot should be able to adapt to that situation. The AI won't have officially conquered dota unless it can actually adapt to any eventuality in the game and this shadowfiend bot didn't come close to that.

    Also btw, openAI is supposed to be an open source team. Their missions statement is to make any advancements in AI open to the public and to developers around the world. Its pretty annoying that they haven't made this bot opensource.
    Last edited by abeltensor; 08-13-2017 at 10:53 AM.

  9. #19
    Quote Originally Posted by abeltensor View Post
    I would like to see the source code as well, Im fairly confident that they used Universe which is OpenAI's neural network toolkit. It would not be surprising if they create a 5v5 full team of bots by next year considering the achievements of google's AI alphaGo which beat professional Go players and then OpenAIs poker bot which destroyed professional poker players.

    It was extremely disappointing that this bot only played a hero like shadowfiend and the player only could choose to play shadowfiend. The shadowfiend vs shadowfiend matchup basically ends after the first 30 secs because of the soul mechanic and as soon as a person dies; the lead against them becomes so overwhelming. I have worked with Tensorflow and Universe and built some cool AIs for different tasks, games and programs. I'm sure a dota bot wouldn't be hard to actually make with a team of 10 experienced developers but putting so many constraints of the players just seems like it defeats the purpose of making a bot like this.

    The bot should have knowledge of all of the heroes and items. It should be able to do more then just last hit, deny, cast raze and kill. For instance, if during a 1v1 match the player chooses to go bot instead of mid, the bot should be able to adapt to that situation. The AI won't have officially conquered dota unless it can actually adapt to any eventuality in the game and this shadowfiend bot didn't come close to that.

    Also btw, openAI is supposed to be an open source team. Their missions statement is to make any advancements in AI open to the public and to developers around the world. Its pretty annoying that they haven't made this bot opensource.
    For their 5v5 bot my guess is that IF they write it, its core algorithm will probably be from this paper. Is this algorithm good for some applications? Yes, definately! Is it good enough for learning to play pro 5v5 Dota 2 in a feasible time frame? I don't think it will be even close! (proposition 1 should give you some intution why)

  10. #20
    maybe it did use api?

    https://twitter.com/mark_riedl/statu...81792901398529

    maybe it can infer what the creeps target should be based on actions.

    I know that when im within 500? range of creeps and i right click a hero theyll now be attacking me. then after 2? seconds theyll no longer be attacking me.

    (some other random stuff https://news.ycombinator.com/item?id=15001521)
    https://github.com/ThePianoDentist/t...dentistdotabot Lina bot which pulls small camp when 'laning' (Aim to work on pulling and stacking bots initially)
    https://github.com/ThePianoDentist/dotabots-ml-tools Parsing data from bot games

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •