back to article OpenAI bots smashed in their first clash against human Dota 2 pros

In the past hour, OpenAI's artificially intelligent bots lost their first match against professional players at smash-hit computer game Dota 2 at The International – the video game's annual championship tournament. It's the first bout in a best-of-three competition between human professional players versus OpenAI's code, the …

  1. anoncow

    Smashed, really?

    It was actually very back and forth until relatively late in the game. "Smashed" grossly mis-characterizes what actually happened.

  2. p3ngwin1

    why surprised humans beat a handicapped AI ?

    "Reaction time for the agents has increased from 80 milliseconds to 200 milliseconds to give the humans players a chance."

    So the humans won because they had a handicap bonus by slowing down the Ai by 150% ??

    The humans are allowed to be their best, why isn't the AI allowed to be its best ?

    I can run as fast, or faster, than Usain Bolt, in the 100m... if you give me a 50 meter start :)

    1. sabroni Silver badge

      Re: why surprised humans beat a handicapped AI ?

      "There are simply too many combinations and possibilities for a computer to master at the moment."

      The article states that there are a lot of restrictions in place because the AI can't handle the full game. Surely letting the humans be their best would involve these restrictions also being removed?

    2. jmch Silver badge

      Re: why surprised humans beat a handicapped AI ?

      "The humans are allowed to be their best, why isn't the AI allowed to be its best ?

      I can run as fast, or faster, than Usain Bolt, in the 100m... if you give me a 50 meter start :)"

      Not a good analogy. Human reaction time is limited by physiological processes, and the physical limit for a single reaction event with full concentration (ie olympic sprinters reacting to starters gun) is about 120ms. 200ms is probably around right for sustained reactions over a long period. So this limit is removing an unfair advantage from the AI rather than giving one to the humans. Besides the AI has other advantages, eg limited game complexity (like simplifying a game for a child), and being able to see the full map. Maybe you could make things 'fairer' by reducing AI reaction time to 150ms and limiting it's game view to anything the agents can see while removing all complexity limits from the game.

      Incidentally Bolt's top speed registered 12.4m/s, and a 100m record of 9.58 sec. Top sprinters can do 100m in sub-10 sec. Top athletes like elite football players can do a 100m in sub-12 seconds, the fastest in around 11. If you're pretty fit and quite fast, you could probably do it in 13-14. If you're like the majority of guys I know in IT, anything from 16-17 to grinding to a halt gasping for breath after 30-40m. Also from a standing start, 50m is not half the 100m time, it's rather more. So I'd give Bolt decent odds even giving you a 50m start :)

      1. rjstua

        Re: why surprised humans beat a handicapped AI ?

        Now that's a race I'd like to see! Usain Bolt vs Ordinary folk with a 50m headstart!

    3. LewisRage

      Re: why surprised humans beat a handicapped AI ?

      It's a fair point, but the AI also gets other advantages, being able to 'see' the map from the beginning where the humans can't, and character selection is limited.

    4. Dave 126 Silver badge

      Re: why surprised humans beat a handicapped AI ?

      > The humans are allowed to be their best, why isn't the AI allowed to be its best ?

      The areas in which the AI was handicapped weren't the areas where the AI let itself down. Irrespective of speed, the AI was making some unforced tactical errors seemingly unrelated to the restraints placed upon it.

    5. georgezilla Silver badge

      Re: why surprised humans beat a handicapped AI ?

      A quick search says that the average human response time is about 200 -250ms. So raising it to 200 for the AI isn't unreasonable.

      Using your logic, why not lower the AIs time as close to zero as you can, so that they could be at their "best"?

      Could you run as fast as a bullet (or faster), even if you were given a 99 meter head start? I mean why handicap the bullet?

    6. streaky

      Re: why surprised humans beat a handicapped AI ?

      why surprised humans beat a handicapped AI ?

      Yep. I've been into Dota for many years and AI holds day job interest for me and I keep telling people that the bots are on easy mode right now and they're giving the best teams on the planet a good run for their money already.

      By the way the more heroes in dota you learn the less complex the learning process. There's also comments I'd make about the AI having access to illusions and heroes that summon things you can micro and it being fa easier for the bots to do things with. I keep having nightmares about formation-flown treants and spiderlings.

  3. Shane 4

    Always remember the "Yet" part when it comes to A.I. vs humans in anything!

  4. jmch Silver badge

    Training time

    Given that the bots get 180 years of gameplay training in a day, and they still have to limit game complexity for them to compete, either they learn VERY slowly, or else there is some limit beyond which little or no further learning takes place with further training. It would be interesting to see the AI play itself at different training levels to get a feel of how fast it progresses with training, and whether such a limit is apparent.

    Can a bot with 100 years training hold it's own against one with 200? Or is the bot with 210 years training comfortably beating the one with 200? That's important to know if further improvement can be made through training or if you've reached a limit imposed by how good the algorithm is and how much CPU power you can throw at it.

    1. Dave 126 Silver badge

      Re: Training time

      My assumption is that more experience might lead to diminishing returns. I'm assuming there's more common moves at the beginning of the game, and rarer scenarios in the middle. After X years of experience, the more common moves may be more or less mastered with little room left for improvement.

      Of course with Self Play, practice doesn't mean having to start a new match each time. The state of a single game could be resumed from half way through many times.

      I'm not a DotA player though - can anyone connector t on how far off I am?

      1. Anonymous Coward
        Anonymous Coward

        Re: Training time

        My assumption is that more experience might lead to diminishing returns. I'm assuming there's more common moves at the beginning of the game, and rarer scenarios in the middle. After X years of experience, the more common moves may be more or less mastered with little room left for improvement.

        can anyone connector t on how far off I am?

        Not very far, or more so right on the spot.

        The game start off with gold collecting as the "common moves". Gold is used for buying items that increase teammate power and overall game advantage. (For gamers, it's called last-hit, where last hitting the periodically spawn minions can give you gold)

        Then rarer scenarios like team battle in the middle. This is where both team grab their weapons and clash together. (For gamers, it's called team-fight.)

        After X years of experience, the more common moves does more or less mastered with little room left for improvement. Frankly, the AI should be able to master this quicker than human by default, especially when there are game cheats available that runs an algorithm to do this for the human players.

    2. streaky

      Re: Training time

      Given that the bots get 180 years of gameplay training in a day, and they still have to limit game complexity for them to compete, either they learn VERY slowly, or else there is some limit beyond which little or no further learning takes place with further training

      It's because they're playing mostly against themselves so they don't really see the intuitive play that humans do, especially some of the cheese strats. When they do see this stuff they learn from it very quickly though.

      Can a bot with 100 years training hold it's own against one with 200? Or is the bot with 210 years training comfortably beating the one with 200?

      I would *assume* that was implicit. Difficult to test if the rules for the bots are constantly changing though - my guess is that there's an element of acceptance testing that it doesn't accept core strategy changes unless it can provably beat the previous strategy algo though, that's how I'd do it anyway.

  5. Pascal Monett Silver badge

    The humans won because strategy - no surprise there

    I have played my fair share of Battlefield and other multiplayer mayhem fests, and if there is one thing I learned it is that the team that has a strategy is the team that wins.

    Most humans go into these games thinking and acting as if they were alone. If everyone plays like that (most common case), then the win is just chance to the team with the more efficient killers. But play a team that has no cohesion and find that the other side has a group that are actually playing together, and they will wipe the map with your corpse.

    Had these tests been against the regular human bozo, humans would have been squashed, no question. It is good that they play these matches against human players who do have team cohesion. That way we see that programming strategy into a statistical analysis machine is not an easy thing to do.

    I wish those programmers all the luck. Or do I ? Did I just hear the whine of a phased plasma ri<no carrier>

    1. Boothy

      Re: The humans won because strategy - no surprise there

      Many years ago I worked in a large office block in the UK, (about 2,000 people), and we used to arrange LAN parties there about once a month. Playing things like UT (Unreal Tournament), and later Counter Strike. (so would have been about 1999 to 2001).

      We'd set up a network in two meeting rooms, one team per room, with everyone lugging in desktops from home (we'd usually borrow monitors from elsewhere in the building, most people still using CRTs back then!).

      After several months one of our guys mentioned he had some friends that did the same at their company, who were about 30 miles away, so we arranged a few us vs them parties. Some at our place, some at theirs (so much easier back then to get guests in past security!).

      It turned out they had absolutely no team strategy at all, they all played as a lone wolf. Not saying some of them didn't have skills, 1-to-1 they were quite good, just no teamwork, at all!

      Needles to say, we absolutely thrashed them on every map, every type of team game. I think in two LAN parties, which would last about 7 hours, they never won a single game. In fact things got heated at one point, where they started to accuse us of cheating!

      So we got one of their guys to stand in our room to watch the next match. while one of ours went into their room.

      Their guy got to see how we played. We were organised into sub groups, squads if you will, and all squad members sat together, and we communicate with each other, and the other squads. Some doing recon, others doing fake attacks to distract. The key thing is we went in with a game plan, and we were coordinated.

      Our guy noticed their room was a complete free for all, no one giving orders, no game plan, nothing. So he started telling them things. "Right, they'll send one squad that way, but it's only a distraction, the real attack will be over there instead." and other general advice, like don't go off on your own, at least be in pairs etc.

      They still didn't win, but at least they now understood why we did.

      They were a much tougher team to beat a month later, turned out they'd been practicing team work in the mean time :-) It was also much more fun for us, as we then had to start upping our game.

      Happy days...

      1. asdf

        Re: The humans won because strategy - no surprise there

        >The key thing is we went in with a game plan, and we were coordinated.

        For whatever reason this story makes me think of Emmet in the Lego Movie basically saying the exact same thing lol.

    2. JDX Gold badge

      Re: we see that programming strategy into a statistical analysis machine is not an easy thing to do.

      They don't program the strategy. They let it learn how to do strategy on its own.

      1. amanfromMars 1 Silver badge

        Re: we see that programming strategy into a statistical analysis machine is not an easy thing to do.

        Here's a Strategy Programming Further Ahead for New Virgin Arrivals .... in Live Operational Virtual Environments Pioneering with the Fab Fabless Mad Delivering Special Rations for Radical Good Guaranteed Supply of Revised Product for Current Presentation ...... Immaculate Unwrapping of AI Perfecting Programming for Star Leading Generations to AIMaster Practical Leverage with Futures Planted in Reality ..... to Grow and Bloom, Seed and Blossom. ........ https://forums.theregister.co.uk/forum/1/2018/08/22/ietf_draft_proposes_encrypted_message_security_for_all/#c_3594445

        And that's Alien Virtual AIMachinery Immaculately Rooted Deeply and Stealthily Routing NEUKlearer HyperRadioProACTive IT Core Source Directions to Extant SCADA Systems for Assimilation via Assumptive Presumptive Projects.

        How else would you make such Stellar Progress in the Search for COSMIC Sources with So Many Other Global Operating Devices laying False Trails with Faked Tales Receiving So Much Undue Attention. It's a Simply Complex AIDiversion of the Office of Space that Registers Times and Presents Futures for Real Enough to be Believed Possible Creation .... with Future AIMentoring of Virtualised Systems Monitoring and Maintenance.... and of no true consequence.

        That's a Full No Quibble Warranty Guarantee Delivering Gross Negligence Losses a Stealthy Beta Preview Partnership which Present Revised and Reinformed Versions of Earlier Provings and Extremely Troublesome Realities with Hardly New for Always Perfecting Correcting Perception for ....... well, has any Advance been made on IT being of Heavenly Sees?

        What says the Program of the Presumption? Do you Agree with the Posit and Premise and Admit ExtraTerrestrial Forces are Targeting Critical Infrastructure and Strategic Tactical Sources for Systemic Flaws/Root Infections to be Future Treated or Completely Comprehensively Destroyed/Obliterated from Live Memory.

      2. Spoonsinger

        Re: we see that programming strategy into a statistical analysis machine is not an easy thing to do.

        Then the AI got bored and went off to make a cuppa tea and invent a cure for cancer - but didn't tell anybody because...

  6. Claptrap314 Silver badge

    Knowledge is power

    The ability to see the entire map--and to parse the critical elements from it trivially--is a HUGE advantage. No was would I agree to play an AI under such a lop-sided deal. Would anyone here play a 10 year-old gamer with such a handicap?

    These marketing stunts are just embarrassing.

    1. amanfromMars 1 Silver badge

      Re: Knowledge is Power and Energises IT Applications ‽ .

      These marketing stunts are just embarrassing. .... Claptrap314

      And also most revealing of the present current failed default state of National Security AI Art, Claptrap314. And that is far too an attractive vulnerability to not exploit fully to rabid excess and extravagant success.

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