4 ms·

OpenAI’s Dota 2 AI steamrolls world champion team with back-to-back victories

chillacy 8y ago
It was only August of last year (2018) when OpenAI barely lost to some good human players, and now here we are. It's starting to feel like AlphaGo all over again. If it does go that route, OpenAI Five will continue to get better as it keeps playing itself, till it leaves humanity behind. Chart of appx AI MMR over time: https://openai.com/five/ https://openai.com/five/
- Kalavala93 7y ago
  Highly unlikely.
natmaka 8y ago
Didn't D. Lenat's 'Automated Mathematician' ('AM') software (from the 'Eurisko' line) perform similarly (let's say 'learning new heuristics by discovery, through training') when it won at a championship built upon a game named 'Classic Traveller', in 1981 and 1982? I can't find any occurrence of the words 'eurisko' nor 'lenat' in the openai.com website. It OpenAI's software ran autonomously, starting with a completely generic version (not taylored for games, especially for Dota) then interacting with humans exclusively through the Dota game (no source code nor parameters modification) it may be a breakthrough. https://aliciapatterson.org/stories/eurisko-computer-mind-its-own https://aliciapatterson.org/stories/eurisko-computer-mind-it... http://users.cs.northwestern.edu/~mek802/papers/not-mine/Lenat_EURISKO.pdf http://users.cs.northwestern.edu/~mek802/papers/not-mine/Len... https://www.lesswrong.com/posts/Myr7PLikhzYgPFhuy/replicating-douglas-lenat-s-traveller-tcs-win-with-publicly https://www.lesswrong.com/posts/Myr7PLikhzYgPFhuy/replicatin...
exabrial 8y ago
Still not playing with normal rules yet. Baby steps I guess.
- thomasahle 8y ago
  Which rules does it still need?
  - dominicwllmsn 8y ago
    I believe that rather than gathering information via pixels OpenAI gets it through the API. As a result it knows exactly where its teammates are at all times, their conditions etc. without any delay time (to my knowledge).
    - thomasahle 8y ago
      Sure, but that's not the same as not playing by the rules. Previous incarnations didn't have invisible units, drafting, wards and more. Hence humans had to adapt to an essentially new game. This version of OpenAI Five appears to play standard DOTA.
      - buzzerbetrayed 8y ago
        It currently plays with the following limitations: - only uses a pool of 17 heroes, instead of the 117 (iirc) heroes in regular Dota 2. - no illusions. Meaning illusion runes don’t spawn, you can’t pick heroes that have illusions, and you can’t buy items that create illusions. So it still isn’t playing full Dota. Definitely getting closer though.
        Scrantonicity 8y ago
        Do they still have invulnerable couriers or was that restriction removed?
        thomasahle 8y ago
        In the first game the OpenAI courier was repeatedly getting sniped, so it seems that's now normal.
      - 8y ago

It was only August of last year (2018) when OpenAI barely lost to some good human players, and now here we are. It's starting to feel like AlphaGo all over again. If it does go that route, OpenAI Five will continue to get better as it keeps playing itself, till it leaves humanity behind. Chart of appx AI MMR over time: https://openai.com/five/ https://openai.com/five/

Kalavala93 7y ago
Highly unlikely.

natmaka 8y ago

Didn't D. Lenat's 'Automated Mathematician' ('AM') software (from the 'Eurisko' line) perform similarly (let's say 'learning new heuristics by discovery, through training') when it won at a championship built upon a game named 'Classic Traveller', in 1981 and 1982? I can't find any occurrence of the words 'eurisko' nor 'lenat' in the openai.com website. It OpenAI's software ran autonomously, starting with a completely generic version (not taylored for games, especially for Dota) then interacting with humans exclusively through the Dota game (no source code nor parameters modification) it may be a breakthrough. https://aliciapatterson.org/stories/eurisko-computer-mind-its-own https://aliciapatterson.org/stories/eurisko-computer-mind-it... http://users.cs.northwestern.edu/~mek802/papers/not-mine/Lenat_EURISKO.pdf http://users.cs.northwestern.edu/~mek802/papers/not-mine/Len... https://www.lesswrong.com/posts/Myr7PLikhzYgPFhuy/replicating-douglas-lenat-s-traveller-tcs-win-with-publicly https://www.lesswrong.com/posts/Myr7PLikhzYgPFhuy/replicatin...

exabrial 8y ago

Still not playing with normal rules yet. Baby steps I guess.

thomasahle 8y ago
Which rules does it still need?
- dominicwllmsn 8y ago
  I believe that rather than gathering information via pixels OpenAI gets it through the API. As a result it knows exactly where its teammates are at all times, their conditions etc. without any delay time (to my knowledge).
  - thomasahle 8y ago
    Sure, but that's not the same as not playing by the rules. Previous incarnations didn't have invisible units, drafting, wards and more. Hence humans had to adapt to an essentially new game. This version of OpenAI Five appears to play standard DOTA.
    - buzzerbetrayed 8y ago
      It currently plays with the following limitations: - only uses a pool of 17 heroes, instead of the 117 (iirc) heroes in regular Dota 2. - no illusions. Meaning illusion runes don’t spawn, you can’t pick heroes that have illusions, and you can’t buy items that create illusions. So it still isn’t playing full Dota. Definitely getting closer though.
      - Scrantonicity 8y ago
        Do they still have invulnerable couriers or was that restriction removed?
        thomasahle 8y ago
        In the first game the OpenAI courier was repeatedly getting sniped, so it seems that's now normal.
    - 8y ago