Two Minute Papers - 2018-08-05
Pick up cool perks on our Patreon page: https://www.patreon.com/TwoMinutePapers The paper "Human-level performance in first-person multiplayer games with population-based deep reinforcement learning" and its corresponding blog post are available here: 1. https://arxiv.org/abs/1807.01281 2. https://deepmind.com/blog/capture-the-flag/ Crypto and PayPal links are available below. Thank you very much for your generous support! Bitcoin: 13hhmJnLEzwXgmgJN7RB6bWVdT7WkrFAHh PayPal: https://www.paypal.me/TwoMinutePapers Ethereum: 0x002BB163DfE89B7aD0712846F1a1E53ba6136b5A LTC: LM8AUh5bGcNgzq6HaV1jeaJrFvmKxxgiXg We would like to thank our generous Patreon supporters who make Two Minute Papers possible: 313V, Andrew Melnychuk, Angelos Evripiotis, Brian Gilman, Christian Ahlin, Christoph Jadanowski, Dennis Abts, Emmanuel, Eric Haddad, Esa Turkulainen, Geronimo Moralez, Kjartan Olason, Lorin Atzberger, Marten Rauschenberg, Michael Albrecht, Michael Jensen, Morten Punnerud Engelstad, Nader Shakerin, Owen Skarpness, Rafael Harutyuynyan, Raul Araújo da Silva, Rob Rowe, Robin Graham, Ryan Monsurate, Shawn Azman, Steef, Steve Messina, Sunil Kim, Thomas Krcmar, Torsten Reil, Zach Boldyga. https://www.patreon.com/TwoMinutePapers Crypto and PayPal links are available below. Thank you very much for your generous support! Bitcoin: 13hhmJnLEzwXgmgJN7RB6bWVdT7WkrFAHh PayPal: https://www.paypal.me/TwoMinutePapers Ethereum: 0x002BB163DfE89B7aD0712846F1a1E53ba6136b5A LTC: LM8AUh5bGcNgzq6HaV1jeaJrFvmKxxgiXg Two Minute Papers Merch: US: http://twominutepapers.com/ EU/Worldwide: https://shop.spreadshirt.net/TwoMinutePapers/ Splash screen/thumbnail design: Felícia Fehér - http://felicia.hu Károly Zsolnai-Fehér's links: Facebook: https://www.facebook.com/TwoMinutePapers/ Twitter: https://twitter.com/karoly_zsolnai Web: https://cg.tuwien.ac.at/~zsolnai/ #DeepMind #Quake #Quake3
Karoly....I have something to admit. I'm not a scholar.
Dumb Jason
shhhh just keep pretending...
fake it until you make it
Name checks out
you're obviously suffering from impostor syndrome. now get, back to work!
I was really hoping to see an AI that mastered bunny hopping and item control in standard duel mode.
Nothing takes long in AI world, haha...
@K3NatCSS were you talking about the Hopper Bot mod for Quake 3?
FPS are far from beeing the most complex and hardest to learn games for AI. Dota was.
5on5, endless possibilities, items, heroes, timings of map objectives, cooldowns, HP/Mana(pool/regeneration).
@Januschka Steinuschka fighting games too
@Januschka Steinuschka I don't think so, we must have numbers to say that, fps have too much considerations, looking at it looks not complex, but you must consider that fps tries to approximate real battles and modern fps have most of that and old fps have the basics of that, and that fps have unpredictable physics composition to make its complexity compared to RTS that have exact maths for it complexity, its hard to say which is more complex, and real battles consider even wind velocity changing in sound, speed, acceleration of the units through an area (like tanks, soldier average speed, cars, supply and etc), geometry (angle of shotting, height descent per time, speed of bullet, map area, map best direction strategy per obstacle, amount of time to wait per obstacle, change in acceleration per obstacle), randomness (bullet with direction changed by wind, prediction of target next position in z, x, y, stray bullets, predict enemy view focus, predict opposite forces coordinates, predict opposite forces z-coordinate since being above or below your ground level is advantageous, predict hiding spots, predicting hidden spots from snipers, grenade final coordinate, mortars accuracy, rockets accuracy and other projectiles accuracy, enemies formation and squads and its tactics and strategies, and etc...), gravity, forces, i could go on and on in how complex is fps, z-coordinate is a pain, imagine you are on a hallway and the enemies know your z position, but you don't know if they are below or above ground level, how you deal with that? I know lot of tactics for that... I play fps since kid and fps have abstract composition of strategy and people miss those, but its pretty complex, also, most of first computer work in the beginning of its history was used by military.
cracks open a monster
yep, quake was a good game
Quake Champions is never going to be as good as Quake 3 or Quake Live because it has stupid abilities and other crap that makes the game unbalanced and dumbed down. That's not what Quake is about. Diabotical will be the true Quake 3 successor.
Quake Champions is never going to be as good as Q3 or QLive because it has stupid abilities and other stuff that makes the game unbalanced and dumbed down. That's not what Quake is about. Diabotical will be the true Q3 successor.
@jona QC is actually pretty good, you still need to learn how to rocket jump and strafe jump, the abilities and diffrent HP and speed just add another layer to the game.
Y'all should also check out QCDE, a fan demake.
@jona If it has all of Q3's mechanics and added "stupid abilities and crap".. how can it be "dumbed down"? If aynthing it has added complexity. You are making no sense..
I love how you are able to instill suspense when narrating AI results.
That is how you know someone is truly passionate about something... I'll admit that this AI hype is making me addict and i understand his emotions, hahaha...
Really exciting because of the sparse internal rewards and long term planning. A step towards AI agents that are useful in real life.
You mean AI agents that CAN OUTCLASS humans in real life... that open pretty scary scenarios if you think a bit further than the simple "ooh it can be useful if I don't find my way home after work" ... think about an enemy squad of robot soldier equipped with that AI and with the movement skills of today's darpa robots... without talking of military autonomous swarm killer drones !!! It is another step toward the abyss. WE MUST STOP CREATING SO ADVANCED AI OR THERE WILL NOT BE LIMITS TO ITS USAGE!
But sex robots
we must stop
you can't be older than 30 years ; )
Linus Heinz Hirschi, yeah ... a lot of girls think that when I get up in the morning... the day after! ;) (don't tell them it's not true!)
You see, losing hope is the first step to reach ... failure!
This sounds like a big jump. It's all happening so fast, it's hard to believe. Good thing I found this channel. You just can't keep up with the speed of progress otherwise.
0:41 Red ball at the bottom spinning in a circle "AI has to learn general concepts"
They are spawn camping blue lol
hold onto YER PAPEEEERS!!
Otakutaru I'll see you neaehxxxtt tiiahhm !
Imagine walking by google headqarters, "This is our FTW!! section ..."
This is incredible work! These developments, for example AI developing its own definition of a reward, are exactly what are needed for more complex AI agents. I was thinking the other day about how rewards would be defined for complex AI problems, and realized that humans can use context to understand whether or not they are succeeding, and use this to determine the actions they need to take to do so. I am looking forward to what this research leads to!
Or the introduction of long/short term memory (maybe it's not new, I just subscribed)
There are a few different memory models that can be used, an LSTM is one option, but i guess it depends on the task.
I don't understand. This doesn't look like Quake 3 at all.
I don't think it would have mattered with this group of human players, they were so predictable I never thought they looked good at any point. I mean if you are playing an AI surely being unpredictable is the first key tactic, changing your tactics every few minutes would be second, then third everybody sticking together as a team and all communicating to target the same enemy at once.
They didn't even look like they were thinking, they were playing dumb computer tactics versus a smart computer lol no chance of a win like that!
If your list of conditions had been applied to the game AND the players weren't absolute dog tripe, then I think the result would have been the other way around. Haha poor dears they got smacked red bottoms, beat by upgraded aim bots.
I've beat aimbots before in Quake, even in a 1on1 match. There are videos of the best players beating cheaters as well. I doubt this AI would win against a team like team liquid in Quake Champions. Probably worse in a 1on1 match as well.
Exactly my thoughts! Any reasonably skilled human team could have beaten this AI, I didn't think it looked at all impressive; the problem was it was not tested as the human players were garbage in the video.
Liquid would absolutely smash this AI, no doubt about it after having seen them at QuakeCon.
Even if they did end up losing a few matches on the trot, you know they would come back with a better strategy and win after that. The only way an AI would be better, is by cutting the oxygen to the room ala HAL from 2001 :D
"Dave Lister" legendary namesake
@pixelasm he said the y still won with reduced aim accuracy.
Hold on to your papers lol
This is insanity!
Ice cream for my eyes.
I love to see AI applied to games! You inspire me to make my own AI videos!
- Congratulations to your team are in order, General, we finally trained the AI guided robot to play paintball games. We even gave her a name, Pandora.
Scientist at 1:15 be like:
We coded that AI too well, it keeps pwning me
Can someone train the AI to trash-talk? Can't wait for that!
+Smilecythe
Quake 3 bots also played Quake 3 18 years ago.
They did neither of these things very well though.
Microsoft's tai did that 2 years back
I mean the Quake III AI already knows how to trash talk
+Nudge Bott
It also knows how to play CTF already.
Easy. Ohh sir...the Insult simulator already does that
When you Set Difficulty game into expert
Amazing! FTW this stuff is advancing every week! Incredible. At this rate I won't be able to hold onto my papers anymore!
2:22 Yet another proof that only noobs camp.
Yup, Quake 3 was a good game.
"was"? You mean "is"!
)))
I WAS WAITING FOR THIS! Thank you!
You know what would be amazing? What if we trained AI to design atoms such that it could use them to build a universe the most efficiently? And what if it completely replicated the known periodic table of elements... or more...? 😱
I get goosebumps from AI :o so exciting
When the AI wipes out me and all my family I can think 'wow cool, I wonder is this the same neural network they used to beat Quake'
Love it :D
Great.
I've been watching Alpha Zero and Leela Zero playing chess at incredible strength. The playing style is a bit 'humanlike' but so much better. Maybe you can do a report on that?
We need this as an official Quake III Arena mod
Low-key solved consciousness
DeepMind vs cooller now
Let's make it Rapha
the best channel ever!!! please do more practical stuff too! :)
Yea right creating an aimbot definitely helps since when the ai figured out how to aim it's game over
HOLY MEATBALLS OF FIRE 🔥 🔥 🔥
I want to see this work on escape from tarkov
AutoML + Reinforcement Learning = Skynet
Muito, muito legal! Quero ajudar a criar a super inteligência artificial...
pfft. Put it against Cypher or Evil. Or Rapha.
Poor DeepMind... It would learn to lose rather, hehe
Looking at demos, i think that decrease of accuracy and reaction time was not big enough. I think, that AI wins just because it aims and shoots faster than human.
If i'm wrong then creators should publish this game and give a chance to anyone to beat this AI. I bet that AI is still bad at tactics and if it cant get advantage of faster aiming/shooting - it will loose.
@taragnor You really misunderstood. Here to aim they are only given raw pixel input. They need to build the idea of enemy and hitting it and figure out it is a rewarding thing it is not like the bots are getting 3d positions and models geometry. Also if you are familiar with AI you would know what they have acheived here is insane.
@Sayam Qazi Image processing was groundbreaking when it was first done, but this is not the first time such a feat has been accomplished using Artificial Neural Networks. Nowadays plenty of AIs are able to derive information from images, so it's not really some new feat in terms of AI.
@taragnor This is not image processing in its traditional sense. Really you have no idea. I will need about half an hour of face to face to explain to you the difference and how hard this feat is. Since that is not gonna happen I am just giving up and hope you will get it somehow from someone else. (Pardon the grammar. English is not my language)
@Sayam Qazi I'd assume it's similar to what self-driving cars have to do to detect obstacles and pedestrians. It seems to be about the same sort of task roughly, though I could be wrong. I know the fundamentals of neural network AI, but I haven't studied image processing specifically.
@taragnor I am just glad you took my comment positively and I wish we could discuss this more but youtube comments are just a very tedious place for it. Again kudos for keeping the conversation nice. One pointer I can give you though. For self driving cars you train them to segment the scene into dangerous/ safe areas. Which is orders of magnitude simpler. In this game though you just give them the input and tell them to succeed with no information whatsoever so even the idea of what is safe and what is dangerous has to emerge for itself. This is a very crude explanation but hope you get the point.
i require this ai in retail game form immediately. see to it.
Wher to download these bots for classic Quake 3?
I want an AI on my team. Can't wait for top tier FPS games to feature AI as teammates
This very competitive Ai in starcraft 2 would be very interesting . I Would like to see which strategies he would use or create
As far as I know they can't really get AI to do much of anything in Starcraft yet. It's hard enough to get them to mine minerals and construct a base let alone scout and counter an opponent.
I've been scammed.
This ain't two minutes,
Its way more..
ty for the vid, I'm going to attempt reading the paper. really wish my school would teach this. this is a great opener for a machine learning course imo. just swap out the game for your first class project
I used my papers to roll a joint before watching this video...
But ai can't figure out how to vacuum my room or mow the grass because they might hurt humans. Humans will always be standing in the way of their progress.
does it have ray tracing?
The comments here make me realise all this research is lost on the majority of the populace
They should make it for UT 2004 :D OR Halo 1 / 2
Bearable escort missions are within an arm's reach!
K1RTB - 2018-08-05
Training an AI team at a shooter game - I don’t see any way how this could go wrong
TheEsotericZebra - 2019-06-03
best comment ever
kduhtdkzrt - 2019-06-05
Multiplayer only games will be a thing of the past in a few years, I guess. Good riddance.
Januschka Steinuschka - 2019-06-07
@K1RTB
#HEwillNOTdevideUS
Quattro 4 - 2019-09-04
@kduhtdkzrt lol no.
TheCreaterKeygen - 2019-10-29
Aw shit, now the XxXSky_Net_1337XxX is going to 360 noscope and then t-bag us. What a time to be alive!