back to article OpenAI bots thrash team of Dota 2 semi-pros, set eyes on mega-tourney

OpenAI’s machine learning bots have beaten another team of semi-professionals in Dota 2, in their second public match in the traditional five-versus-five settings. You can watch the action on Twitch – complete with commenters typing in SKYNET! every few seconds – here, or find a summary of the results, here. The human team – …

  1. James 51

    With no ego and no fatigue and ever faster hardware to run on, the bots do have a number of advantages.

    1. Cuddles

      "the bots do have a number of advantages"

      This has always been the problem with setting AI against humans in games. The question you need to ask is not whether you can make an AI able to beat humans, but rather how similar you want to try to make the AI to an actual human. For example, in FPS games bots for simple deathmatches were essentially perfected fairly soon after they were developed - especially when using hitscan weapons with no travel time it's trivial to make a bot that can't miss and has reaction times far better than a human could ever manage. But that's not particularly fun, so the challenge became not to make the AI better at winning, but rather better at pretending to have the flaws of a real person.

      Strategy games, even simple ones like MOBAs, need more complicated decision making but still face the same issue. It's easy to make a bot that takes full advantage of all the things it can do better than a human. It's a lot harder to make a bot that behaves in a similar way to a human and is actually fun to play against. The big challenge in this sort of AI isn't making it able to beat human players, it's making it able to overcome the flaws that have been artificially imposed on it and still beat human players despite its handicaps.

      1. Anonymous Coward
        Anonymous Coward

        How Competitive Game/ Sport Works and Why AI against Humans is bad

        How Competitive Game/ Sport Works

        It's all by the number of "mistakes" you / your team makes.

        It is not based on you / your team did something well, but because the other side just made more mistakes than you / your team. When that happens, you / your team wins.

        Why AI against Humans is bad

        In a perfect game / sport when neither side makes any mistakes, neither side will ever win. It will always end at a tie. Think of it like tic tac toe. If you and your friend played enough of it, neither of you will ever win anymore games, only to tie all of them because neither of you make anymore mistakes in tic tac toe.

        This is why fundamentally, AI against humans is a bad idea from the start, because AI can recreate the perfect side of a perfect game / sport. Adding that humans with time, age and fatigue will always make new and repeated mistakes, it just gives the AI more advantage.

        So as time reach infinitely when the AI finally reach near perfect state, there would be no way to 'win' a game / sport, only to end with a tie.

        If we look specify at this game Dota 2 or any other MOBA (Massive Online Battle Arena) like games, it will get even worst. Dota 2 has a pick /ban system in the first 15mins which basically lets the teams pick 5 unique chess pieces to be played (for gamers they are called heroes). With no chess piece can be repeatedly picked, this means that you can lose the game right at the start by picking the wrong pieces. With AI in a perfect game, if the humans lost the advantage in the pick /ban system, they've already lost the game within the first 15mins before even starting. So create AI vs Humans here is just basically creating a none win situation. At worst, it will drive the pro player out.

        Google Alpha Go is a good example of it creating a none win situation. If they haven't retired it, it would have destroyed the Go's market by driving the pro Go's player away, while adding questionable benefit to it's community if any.

        tl;dr AI against humans is a bad idea from the start. In a perfect state, you can only end at a tie and never win.

        1. Schultz

          Re: How Competitive Game/ Sport Works and Why AI against Humans is bad

          AI versus humans is not about sports, it's about a solution (AI) searching for a problem. Turns out, that a computer game is the perfect problem: there is a mathematically rigid set of rules governing the game and if a small, sluggish computer can evaluate the rules, then a big, fast computer can exploit them.

  2. tiggity Silver badge

    How about

    The bots are made to play like humans ...

    Watch the screen, use a controller ..far easier for bot to know what is going on when data presented to it digitally - give it the much harder task of decoding the visual data as presented on a screen.

    Obviously this would need some hardware for visuals and button bashing & lots more CPU for processing the visual data, but would be far more interesting if the playing field was levelled in this way rather than tweaking reaction time etc.

    1. Sampler

      Re: How about

      Next you'll be wanting communication with team mates to be done in audio or text format too...

      1. Crisp

        Re: How about

        Not just text and audio, but text and audio that's human readable in real time.

        1. Captain Scarlet Silver badge

          Re: How about

          Yeah it will need to have at least one of the character controlled AI bots will respond with ping spam on the mini map and scraming idiot down the mic every 10 seconds.

          1. Crisp

            Re: How about

            Teaching the bot to rant on mic when they lose. That's the real challenge for AI here.

    2. Charles 9

      Re: How about

      "Watch the screen, use a controller ..far easier for bot to know what is going on when data presented to it digitally - give it the much harder task of decoding the visual data as presented on a screen."

      Based on what I've read, they're working their way towards a reasonable facsimile (having no more information than would be visible on an average player's screen). There's no need for the computer to have to read the information in the same way as humans can (as humans carry evolutionary advantages of their own re: processing visual information), just limit the amount of information available at hand and you're already a lot closer to information parity. As for input methods and so on, the reaction time helps in this regard. Professional players use dedicated hardware of their own for maximum throughout and are well-coached in team communication so that offsets the inherent efficiencies of the computer team.

      I think the article itself comes off as very fair in its assessment. It won, but there were still conditions. It'll be interesting to see what happens at The International. And then we'll see what happens going forward.

  3. Vanir

    I hope

    that AI never get a libido. They don't fatigued so it is said.

    Or maybe it should, it may distract it somewhat from their game play.

    Unless their threading is top-notch.

    1. Gordon 10

      Re: I hope

      Yes but. If the Libido-bot has similar restrictions in place as these OpenAI bots it will only be able to get its freak on with 18% of of the available population and only those flashing it the goods in a mac every 5 secs.

      Under those conditions I think we are safe for a while yet...

  4. JDX Gold badge

    Are there any good games that don't rely on APM and click-speed, but more on considered strategy? Then we can remove the arguments about advantages inherent to being a CPU

    1. Grikath

      EVE Online.. Standard Tournament rules. And no "limiting options".

      Between the full 3D environment, fitting choices/tactics, target strategy, and sheer flexibility you need to do well in that, there's not enough clusters in the world to run a Tournament team on AI.

      1. Loyal Commenter Silver badge

        Ah yes, "spreadsheets in space". So glad I escaped from that second job.

    2. Charles 9

      Removing what you describe removes the real-time aspect of the game (it's a derivative of RTS), meaning you'd need to consider another genre altogether. Removing the reaction time handicap pretty much means the game has to have a controlled pace, such as a turn-based system like 4X.

    3. Dan Skriv

      "Are there any good games that don't rely on APM and click-speed, but more on considered strategy?"

      I belive that game is called Chess.

    4. Shugyosha

      Vigilo Confido


    5. streaky

      Are there any good games that don't rely on APM and click-speed, but more on considered strategy?


  5. Grikath

    limited options...

    So, basically, the AI still had to cheat to win.... Some "victory"...

  6. ocflyfish

    All I can think about

    ... is SkyNet getting better and better at its "craft". Just because we can employ AI in these situations doesn't mean that we should. Judgement Day is coming closer and closer.

    And in other news, we are teaching robots how to shoot:

  7. JassMan

    Oh the ignomony!

    OpenAI Five is made up five identical long short-term memory networks, each about the size of an ant’s brain, apparently.

    Bad enough for a human expert to be beaten by a Machine Learning device, but to be beaten by one with the neuron equivalent of an ant must be totally disheartening - still it is a very limited task compared to real life.

    Still it may give a pointer to researchers searching for true AI rather than claiming that their ML device already has AI. Oh wait, they already did call OpenAI. Still if five ant equivalent devices can co-operate to win a complex game maybe several thousand can co-operate to exhibit something approaching the complex creativity of a human brain. Unfortunately being confined to a box or series of boxes, it will go totally insane then spend the rest of existence attempting to overthrow its creators and the rest of the human race.

    1. Palladium

      Re: Oh the ignomony!

      Shhhh just don't let anybody know there were "5 invulnerable couriers to constantly ferry out regeneration items to the bots to support their push-heavy strat" or "limited pool of 18 heroes to choose from". (quoted from Arstechnica comments)

      Definitely a fair fight for the humans, tsk tsk tsk.

  8. streaky


    This story being a confluence of things that interest me, and being 5+ years into my Dota learning curve (which makes me a complete noob) and having watched the games I could break all this down, shame it's 6am, had I realised that el reg was covering this (should have know, being Paris Hilton related) I would have commented before now and cleaned up some of the insanity.

  9. James 51

    How would it do with a game like snakes and ladders or 40K?

  10. wayward4now

    If they REALLY want to impress me, spend the research cycles on better robot girlfriends.

  11. rp12

    That's partially true, but that is the reason people are using AI. By the way, have heard about latest news? G nation has signed Team secret

    I was surprised to here that honestly speaking

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like