DeepMind: The Hanabi Card Game Is the Next Frontier for AI Research
❤️ Pick up cool perks on our Patreon page:
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
313V, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Anthony Vdovitchenko, Brian Gilman, Christian Ahlin, Christoph Jadanowski, Claudio Fernandes, Dennis Abts, Eric Haddad, Eric Martel, Evan Breznyik, Geronimo Moralez, Javier Bustamante, John De Witt, Kaiesh Vohra, Kasia Hayden, Kjartan Olason, Levente Szabo, Lorin Atzberger, Marcin Dukaczewski, Marten Rauschenberg, Maurits van Mastrigt, Michael Albrecht, Michael Jensen, Morten Punnerud Engelstad, Nader Shakerin, Owen Campbell-Moore, Owen Skarpness, Raul Araújo da Silva, Richard Reis, Rob Rowe, Robin Graham, Ryan Monsurate, Shawn Azman, Steef, Steve Messina, Sunil Kim, Thomas Krcmar, Torsten Reil, Zach Boldyga, Zach Doty.
Thumbnail background image credit:
The Hanabi card game images are owned by the published and one image by Drentsoft Media –
Poker image:
Splash screen/thumbnail design: Felícia Fehér –
Károly Zsolnai-Fehér’s links:
Facebook:
Twitter:
Web:
#DeepMind #Hanabi #AI
In other words, they'll need to create a general AI that can learn to talk. I would think first try and see what the "grammar" is for this kind of communication, and then see if the AI can use these grammar rules instead of trying to learn the grammar from scratch.
Fun & very much related fact: Hanabi is NP-hard – even for players who manage to cheat and look at their cards! See here:
https://arxiv.org/abs/1603.01911
Citations here:
https://scholar.google.com/scholar?cites=7635630622238913382
Magic: The Gathering (MTG) can be researched and implemented too… There are plenty of rules enforcement engines and AI supports like XMage ( https://github.com/magefree/mage ).
Come over to our hanabi discord channel if you'd like to learn how to play it with the pros! https://discord.gg/wusacX or visit the website at https://hanabi.live/ – RaKXeR
4 years later: tic-tac-toe is the next frontier of AI research.
sc2 is really intresting because zerg is very very not clear
Cool, I recently thought about similar problem. When playing the, a lot of progress comes from practice with other person like coach or friend. Algorithms have something similar with actor critic and GANs, but I never saw approach where networks play in some sort of training mode where they try to find best possible solution or point towards the right move in some situation. Or network that is similar to ones that predict quality of move, but predict surprise factor (what move opponent expects) might be an idea.
Is some work this exists?
Who is up for a little research/project?
awesome!
Well. AI reached superhuman performance at Poker quite some time ago. So no Deepmind, Hanabi is no fronteer at all.
AI sucks at Hanabi vs humans because humans are dirty cheaters, AI doesn't stand a chance.
My first question was "What is Hanabi?"
Sorry to disappoint.
The jokes will keep coming. You asked for it. Very interesting and clearly there is room for improvement. Ai will soon invent their own language. That would be awesome to see.
I feel like this is probably a game you could explain, because I really can't figure out anything about the game from what you said. What degree of communication is allowed between the players?
For people who don't know what Hanabi is: Every turn you can either play one of your cards, throw away one of your cards, or you can give a hint to one other player. For this hint you need to choose either a colour or a number, and then you tell this player which of the cards in his hand are this colour or this number. When you play a card, it must fit on to one of the cards lying on the table. The card fits when the card of the same colour and the number directly below it is on the table. If you play a card that doesn't fit, it counts as a mistake. Three mistakes and it is game over.
So you need to be really clever about which information you give (and also which information you do NOT give), so that the other player can infer which cards they should play or throw away. If you play with more players you see more cards, so you have more information, but since the pool of hints you can give is limited (and replenished by throwing away a card), other difficulties come into play.
I hope they keep working in sc2, they still have a lot to prove.
Fascinating what seemingly simple things are the hardest for computers. I wonder what'll be next
We are getting into imperfect information co-op games now! How exciting!
If we find a good structure for this game, it might be interesting to see how applicable it is to the language processing field, considering it would be optimized for language's main purpose.
– researchers successfully teach ai to cooperate with other ai
– researches teach ai to cooperate with humans
– ai realizes other ai are more valuable teammates and humans are obsolete
– everyone dies
I love this!
Excellent summary and explanation as always. I can tell you spent a lot of time digesting the paper
I was thinking of drawing up an AI for hanabi… haha
they didnt even beat Starcraft 2 yet. All races, all maps, top1 world player.
We are three years away from AI capable of playing all these games with super human levels.
The hardware needs to go up another 100X to make these algorithms go mainstream in the future.
I wonder how well these AIs would perform at Hearthstone?
And I just want them to beat minecraft. Like it's got imperfect information, sparse rewards and also long term planning.
AlphaGo didn't beat the pros fairly.
Next up: AI going online and searching for new games to play
The principles that govern players in Hanabi closely resemble the same ones that govern actors in a military tactical situation. If an AI can nail this then the problem of real time autonomous battle field machines working optimally together is a huge step closer.
I just saw bits and pieces of Googe's Stadia presentation and I'm surprised to already see a realtime application of style transfer in a big product.
What about the card game Bridge?
I hope deepmind never learns to play a game I enjoy
Finally! I've been following AI research for a long time, and I'm also really into boardgames. I've always known that boardgame (type games) are the deepest type of learning as it's the only area which I myself can still grow and learn in. The complexities between working with other free agents (humans in our case) is just unmatchable! Glad they are in a place to finally start cracking these challenges! Also AI research is finally moving into an area I know extensively! This game in particular is perfect as a test bed, I highly recommend people try it out to understand the subtle complexities in it! If you find it too easy, you need to play against better players!!! We're really getting close to true AI now! 🙂
Soution: An AI that instead of trying to play writes a very narrow "stupid" AI which will win the game.
The consensus is that they didn't do anything that impressive with Starcraft. It didn't have the same limitations players did, and it was only able to run a single strategy. It won by performing more actions per second than a human player would ever be able to match. It also performed actions in multiple places on the map at the same time, because it wasn't limited to the same window that human players are. At best it was a gimmick, and was in no way as profound or impressive as the victory at Go.
Oh man your videos are great, but that "mute" audio at the beginning of each new sentence is difficult to bear. I hope your audio software settings get revised on your next videos. Thanks.
Hello person scrolling through the comments
very well explained and concise
köszi Károly! )
You guys heard it first here, in Youtube. DeepMind is done with SC2 scrubs and now it wants blood in a much harder game, in Hanabi.
The weirdest board game no one ever played but somehow got a board game of the year award.
Next in line are, Munchkin, Exploding Kittens, Pathfinder, Rust and Minecraft.
This tells me when DeepMind was playing StarCraft, it had perfect information of the map whereas human players could only see a limited field of view through the screen. Cheater.
Could aversarial neural network work as well where the adversary for player agents is the card shuffler who tries to make sure that as many hands as possible would lead the players into loss? Obviously, there are some combinations that always lead to loss but what about those that aren't?
Game Theory scenarios can get hugely complex to compute. I'm a little scared of what we might interpret from the results of machine learning and game theory.
I can't believe it! I am working now privately for about half a year on the exact same problem. Unfortunately, my neural network does not perform better than 6 points on average which is still quite dumb. My experience is that it is somehow difficult for the AI to get the sweet spot between playing risky and playing super safe. I often found that after training, the AI would often discard cards like 90% of the time even if it does not make any sense just to end the game without making mistakes… The other extreme is playing cards way too early and loosing 90% of the time…
AI in Starcraft II with low APM CANNOT beat human pros!!
Good that you said they made 'progress'!
So now we’re teaching AI to communicate with other AI using imperfect information. I’m studying AI and am all for it, but this is the one that scares me as existential threat.
It's true, though – moving from perfect information games to imperfect information games is a HUGE step forward for AI. There's also an issue inherent to card games that makes them several orders of magnitude harder to cope with than a game like chess – randomness. Every game of chess starts off in exactly the same way. With card games, the possibilities are almost literally infinite.
And collectible card games would be another entire order of magnitude above that…
Possible to simply train 4 different models and make them train together? Would reduce the knowledge each model has of its opponents.
So why not Bridge too ?
I appreciate that you said "Great progress in Starcraft II" instead of having "Beaten" Starcraft II. Alpha Star was deficient by several metrics.
I feel like these videos are always missing some key piece of information to make them truly great. For example, most people are probably looking up the rules to Hanabi after watching this, which means that all the video accomplished was saying "This work exists", which is not very useful.