16

Deep Blue defeated Kasparov in 1997, marking a milestone in which computers triumphed over humans in chess. This event is considered pivotal in the history of humans versus machines in the realm of chess.

Does this "championship" exist in computer games?
Bots integrated into games are balanced to allow humans a chance to win. However, what about the use of AI (or more broadly, computer programs) targeted to defeat humans in the popular competitive multiplayer games (Starcraft, Dota, Counter Strike, etc)?

6
  • 17
    "Bots integrated into games are balanced to allow humans a chance to win." Not always. Sometimes, they are balanced to give humans a challenge by cheating (Warning, TVTropes link).
    – Nolonar
    Feb 20 at 14:14
  • Not a tournament or a competitive game, but you may be interested in Rog-O-Matic. It was an expert system (not really AI, but conceptually similar) built to play Rogue. It pretty consistently outperformed a majority of players, even experts at the game. Feb 20 at 21:54
  • 6
    I don't imagine many would say humans can beat AI at first-person shooters (with hitscan weapons), given that AI has reflexes that are orders of magnitude faster than that of even the most reactive humans, and their aiming is pixel-perfect, and those are huge parts of the skill of first-person shooters. This has been true for decades. The only place that humans can even conceivably hope to compete is with the more strategic or predictive parts of those games, or if the speed and precision of the game maxes out roughly where humans are at.
    – NotThatGuy
    Feb 21 at 10:36
  • On Super Smash Brothers Ultimate, it is quite common - and funny - to see pro players lose to Max Difficulty bots.
    – T. Sar
    Feb 22 at 17:33

6 Answers 6

33

I'm not entirely sure what you mean by "championship", but if you are referring to an instance where an AI beat the world champions of a game, specifically in competitive multiplayer game or esports, then yes there is at least one occurrence of this happening that I am aware of.

Specifically, back in 2019, OpenAI Five beat the then world champions (OGesports) in a best of 3 match in DotA 2 with a 2-0 score line (NOTE: there was a difference in rule set during this match compared to a standard match, including a restriction on heroes played and items able to be purchased). OpenAI Five and DeepMind's AlphaStar had previous wins against other pros in private matches, however this marked the first occurrence of an AI beating a professional esports team in a livestreamed tournament match. You can read more about it here, as well as find links to the gameplay.

As for Starcraft II, DeepMind's AlphaStar is reported to be better than 99.8% of all human players, and

"was able to best top pro players 10 matches in a row during a prerecorded session, but it lost to pro player Grzegorz “MaNa” Komincz in a final match streamed live online".

12
  • 4
    @Keltari I can speak to the Dota 2 side of things: The AI primarily learned by playing itself. It started with 1v1, then scaled things up to 5v5. There's obviously some coding to interface with the game, but beyond that, it learned the strategies (denying, when to go aggressively, etc) all on its own. Feb 21 at 1:31
  • 10
    @Keltari To be honest, I'd be even more impressed if someone directly programmed a software system that plays FPS the way a human would (by interpreting rendered images and sending controller inputs). "AI" just means using machine learning to generate programs that we don't know how to actually write; for the things we use AI for "teaching an AI to do that" is much easier than directly programming it. FPS bots that could be programmed to get instant headshots work by having direct access to the (at least some of) the game state instead of using machine vision to interpret an image.
    – Ben
    Feb 21 at 4:31
  • 7
    I would argue that the DotA 2 example is not all that good. The games played were under different rules than normaly. The OpenAI won by using tactics that just doesn't work in a normal game, and that caught the human team completely off-guard. If we made comparison to chess: it would be just as the players would play without knights, and every 5 turns a player could make 2 moves instead of 1.
    – Negdo
    Feb 21 at 10:01
  • 4
    @Negdo I'm not doubting you, but could you go into the ways that particular DotA match may not have mirrored a "real" match?
    – Michael W.
    Feb 21 at 21:08
  • 6
    @MichaelW. Both teams played under the following restrictions: limited to only 18 heroes (out of >100), certain items were not allowed (bottle, divine rapier), any hero or items producing summons or illusions was also disallowed, as well as scanning, and any sort of courier-based tactics (teams played with 5 invulnerable couriers each). Source: openai.com/blog/openai-five-benchmark Feb 21 at 22:14
19

There're several games not mentioned yet in which AI is capable of playing/beating the best humans in the world (or people are working on it):

  • Go. The AI that first became superhuman for this game trailblazed the entire AI revolution in gaming.
  • Poker, where the bot is superhuman. The bot is considered dangerous enough by its developers that it's not publicly available, since using it to cheat would ruin the online Poker scene.
  • Hanabi has been called the latest frontier in AI research by DeepMind. A hallmark of this game is that it's cooperative as opposed to competitive; another hallmark is that you can see everyone else's cards except your own. Neural-network-based AI seem to be exceptionally poor at this game compared to hand-coded ones.
  • Computers are still pretty inferior to humans at contract bridge (although they did recently beat some top humans in a highly restricted version of bridge).
  • It doesn't look plausible for computers to play Magic: the Gathering as well as humans, much less build Magic: the Gathering decks as well as humans. (This is a collectible card game where you first build a deck and then you play against other people with decks that are not necessarily your own.)
  • As far as I can tell, there are no serious attempts to get computers to play Eleusis. (This is an induction card game where players attempt to guess a "rule" which can be almost anything.)
  • Computers might never beat humans at Calvinball. (The graphic is old, since computers are now better at Starcraft & Poker, but I don't believe there are any serious claims that computers can play Calvinball, let alone beat humans at it.)

Finally, edge case:

  • Chess (!). Chess computers have beaten humans at chess since 1999, but they've not beaten human programmers, since the current-strongest chess engine (Stockfish) uses a lot of human heuristics in its search algorithm. Stockfish is stronger than Leela Chess Zero, which is a self-learning AI for chess. (But note that Stockfish also relies on neural networks for its evaluation these days, so it's sort of a hybrid.)
10
  • 2
    Interestingly like, I think LLMs would be pretty good at Calvinball. There's a whole market of virtual chatbot friends/roleplaying as george washington, right. They'd do fine at "nuh uh! you didn't touch the 6th base, so you can't take the shortcut" arguments
    – Kaia
    Feb 21 at 18:52
  • 2
    @Kaia Hmm I guess with a very high temperature setting, LLMs could conceivably play Calvinball. I'd be curious what they come up with though. Like if I were to say "you stepped into a punishment zone, so you have to sing a song" they'd probably sing a song and then wait for my next prompt - at which point they are hardly winning.
    – Allure
    Feb 22 at 0:41
  • @Allure: The reason Calvinball appears in that comic is that it is so ill-defined that there is no such thing as "winning" or "losing," so computers will never win at it (nor will humans). Note also that Stockfish, Leela, and AlphaZero all appear to be pretty close to one another (for example, Stockfish actually managed to lose a few seasons of TCEC a few years ago).
    – Kevin
    Feb 22 at 6:47
  • @Kevin I feel like that's the point though. Computers need rigid and well-defined rules. Once that's not available, they are no longer able to handle the game well. (Presumably this is also why computers can't play Eleusis.)
    – Allure
    Feb 22 at 7:31
  • 2
    While most have computer adaptations, I don't think any game you mentioned matches the question's criteria. Feb 23 at 7:40
9

Rocket League is actually an interesting frontier in AI-controlled, high-skilled play. The best Rocket League AI a couple years ago was able to beat about 80-90% of players consistently. I have no idea how much they've improved on that by now, but I do think the best AI now is up to around 99% at least - I've heard only professionals in Rocket League can really stand up to the top tier AIs these days.

These bots are trained in a proper AI way, a similar training regimen to how AIs like alphago are trained - they are neural nets, I believe, which are just forced to play each other over and over again and again as they build pathways in their neural net to improve decision making. They're not hardcoded with a set of techniques that work in certain situations - they're proper learning machines, and they learned, over probably hundreds of thousands of hours of play, how to be as good as they are.

Look up videos about Necto and Nexto for reference.

3

There is a chess computer championship on chess.com here:

Computer Chess Championship is a constantly-running (24/7/365) tournament played by the best chess engines in the world...

However AI is not allowed to play with humans in actual human world championships, because they are just too good at the game at this point.

1
  • 1
    I dont think this answers the question, as playing chess isn't against humans as you noted, and playing chess isn't really a "computer game" in the sense of video game.
    – qwr
    Feb 21 at 5:30
0

Go (aka Weiqi / Baduk). Imho this is the most incredible accomplishment of AI. Check out the story of AlphaGo, whose success led to a resurgance of neural network architectures in gaming. With various successors reaching similar strength, such as Leela Zero.

The chess version of which (Leela Chess Zero) twice winning TCEC against Stockfish before it integrated neural network elements itself, since which it hasn't lost TCEC.

-8

There is a Tetris 1v1 championship. They don't allow computers to play. You don't need an AI or even a powerful computer to write a program that will play Tetris perfectly.

Games like Tetris and even chess are very simple. At any given point in time there are only a few moves that can actually be made. Rotate the block, move the block, drop the block. In chess, there are more moves. Move the pawn one square, move the pawn two squares, move the rook forward, move the rook back, capture the knight, etc. There are far more moves, but they are limited by the pieces and the board.

As for true AI on games like StarCraft, DOTA, etc, it is not likely an AI will even be remotely competitive. True AI needs to be trained and tuned, which takes massive amounts of time and resources. At any point in time in a live action game, a near infinite number of decisions can be made. Add in more units/players/maps/whatever, the number of decisions skyrocket exponentially. However, given enough time and resources, an AI should always become unbeatable by a human.

4
  • 11
    You'd be surprised. AI's are already able to beat many professional SC2 players and Dota2 teams.
    – Dragonrage
    Feb 20 at 21:57
  • 10
    Unfortunately the last paragraph is just plain wrong.
    – qwr
    Feb 21 at 5:28
  • Additionally, the number of decisions is exactly as high as the number of "input keys". So much smaller than you make it out to be. Feb 21 at 8:49
  • 3
    @Dragonrage For Dota2 that isn't actually true. That showcase game from a few years ago used rules completely different from normal Dota game.
    – Negdo
    Feb 21 at 10:04

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .