DeepMind's AI Beats Humans At Quake III Arena (yahoo.com) 98
"A team of programmers at a British artificial intelligence company has designed automated 'agents' that taught themselves how to play the seminal first-person shooter Quake III Arena, and became so good they consistently beat human beings," reports AFP:
The work of the researchers from DeepMind, which is owned by Google's parent company Alphabet, was described in a paper published in Science on Thursday and marks the first time the feat has ever been accomplished... "Even after 12 hours of practice, the human game testers were only able to win 25% of games against the agent team," the team wrote. The agents' win-loss ratio remained superior even when their reaction times were artificially slowed down to human levels and when their aiming ability was similarly reduced....
The team did not comment, however, on the AI's potential for future use in military settings. DeepMind has publicly stated in the past that it is committed to never working on any military or surveillance projects, and the word "shoot" does not appear even once in the paper (shooting is instead described as tagging opponents by pointing a laser gadget at them). Moving forward, Jaderberg said his team would like to explore having the agents play in the full version of Quake III Arena and find ways his AI could work on problems outside of computer games. "We use games, like Capture the Flag, as challenging environments to explore general concepts such as planning, strategy and memory, which we believe are essential to the development of algorithms that can be used to help solve real-world problems," he said.
DeepMind's agents "individually played around 450,000 games of capture the flag, the equivalent of roughly four years of experience," reports VentureBeat. But that was enough to make them consistently better than human players, according to Ars Technica. "The only time humans beat a pair of bots was when they were part of a human-bot team, and even then, they typically won only five percent of their matches..."
"Humans' visual abilities made them better snipers. But at close range, [DeepMind's team FTW] excelled in combat, in part because its reaction time was half that of a human's, and in part because its accuracy was 80 percent compared to the humans' 50 percent."
The team did not comment, however, on the AI's potential for future use in military settings. DeepMind has publicly stated in the past that it is committed to never working on any military or surveillance projects, and the word "shoot" does not appear even once in the paper (shooting is instead described as tagging opponents by pointing a laser gadget at them). Moving forward, Jaderberg said his team would like to explore having the agents play in the full version of Quake III Arena and find ways his AI could work on problems outside of computer games. "We use games, like Capture the Flag, as challenging environments to explore general concepts such as planning, strategy and memory, which we believe are essential to the development of algorithms that can be used to help solve real-world problems," he said.
DeepMind's agents "individually played around 450,000 games of capture the flag, the equivalent of roughly four years of experience," reports VentureBeat. But that was enough to make them consistently better than human players, according to Ars Technica. "The only time humans beat a pair of bots was when they were part of a human-bot team, and even then, they typically won only five percent of their matches..."
"Humans' visual abilities made them better snipers. But at close range, [DeepMind's team FTW] excelled in combat, in part because its reaction time was half that of a human's, and in part because its accuracy was 80 percent compared to the humans' 50 percent."
Why, why, why? (Score:5, Interesting)
Re: Why, why, why? (Score:1)
So it can create synergies collaboratively training other skynets in a forward looking, open environment, pushing into new realms, and creating residual income opportunities and new market positions, while, at the same time, using sustainable techniques to face challenges head on and climb to new plateaus.
Seems obvious. (Score:3)
Clearly we're maximizing the odds for destroying humanity. No need to put all our eggs in one basket and hope that war or climate change does humanity in. Earth is prime real estate and it's wasted on those hairless apes but the law is still the law: no invading planets with beings capable of higher thought. #AngelInvestor #ImNotAnAlienYourAnAlien ;)
Re: (Score:2)
Reminds me of this:
Re: (Score:2)
Re: (Score:1)
For the same reason people kept building more powerful nukes I guess.
Re: (Score:2)
We are not doing that. You are hysterical and ignorant of the facts. The automaton described here has zero understanding and zero insight in what it does. It would have lost badly if the "contest" had not been rigged in an extreme way by honor-less and despicable liars.
Re: (Score:2)
https://youtu.be/8JtnEUPvpus [youtu.be]
Re: (Score:2)
AI good at twitch shooting (Score:2)
This was news about 20 years ago?
Esports was fun while it lasted (Score:2)
Esports was on an exponential growth curve to exceeding Hollywood and sports revenues. Now what? Who wants to watch Bender play twitch. Well wait, that's a bad example. I'd definitely pay to watch drunk snarky robots play games. But not an Nvidia card play a game.
Esports Market crash in 3 years.
Re: (Score:2)
Yeah, I reckon I could solo the superbowl if I'm allowed a Challenger 2 tank on my team.
Re:AI good at twitch shooting (Score:5, Interesting)
This was news about 20 years ago?
I too remember getting wrecked by bots in the Quake 2 days. The article belaboring how much better the AI was than the humans is misleading, that's not the achievement. The achievement is due to *how* these bots learned to play. They didn't have paths pre-programmed by a script, they learned everything about the game and the maps organically including the concept of capturing the flag. Where the bots of yore were probably big ass scripts defining all of their behavior with fuzzy logic and what not, these are using more sophisticated methods to achieve a similar end-result but without someone engineering it specifically to accomplish this goal.
Re: (Score:3)
Re:AI good at twitch shooting (Score:4, Informative)
In fact if anything I'm sure history will show that building up a representation of the world from imagery was a much more difficult problem than the team tactics themselves. We just take it for granted because it's on our hardware (wetware).
Doesn't mean the AI is good (Score:3)
Re: (Score:3)
"Humans' visual abilities made them better snipers. But at close range, [DeepMind's team FTW] excelled in combat, in part because its reaction time was half that of a human's, and in part because its accuracy was 80 percent compared to the humans' 50 percent."
Wow, they made a computer with faster reaction times than a human. What next, a computer that can calculate projectile trajectories faster than humans? Von Neumann would be proud (sarcasm).
Re: (Score:3, Insightful)
It would be impressive if it actually was robotic input of normal input devices combined with machine vision that used a camera pointed at a monitor.
Re: (Score:2)
Re: (Score:2)
Robots aren't going to interact with the world by manipulating inputs designed for people.
Indeed not, but androids will.
No, it does not (Score:2)
It does when the conditions are extremely skewed in favor of the automaton. The way this is done is the same, deeply dishonest and dishonorable thing Google did and others as well: The Automaton gets to see tons of games played by the humans and the humans get to see zero games by the automaton in advance. Also, the humans are given no time to experiment and adjust to the way the automaton plays. As the automaton has zero intelligence and absolutely no understanding regarding what it does, the humans would
Re: (Score:2)
Indeed.
Re: (Score:2)
You are an idiot. If things are set up unfairly enough, a random mover combined with an aimbot would win. There is absolutely nothing to be amazed of in that approach.
Why to I get the impression that Google, Microsoft and others have paid shills on ./ and all are posting as ACs?
Re:No, it does not (Score:4, Interesting)
The way this is done is the same, deeply dishonest and dishonorable thing Google did and others as well: The Automaton gets to see tons of games played by the humans and the humans get to see zero games by the automaton in advance. Also, the humans are given no time to experiment and adjust to the way the automaton plays.
Note that when the Starcraft human was given time to reflect after playing against the AI, the human won and the AI lost rather badly.
Re: (Score:2)
Not a surprise.
Re: (Score:1)
> Note that when the Starcraft human was given time to reflect after playing against the AI, the human won and the AI lost rather badly.
You did not provide source for your claim, so I will add it:
https://www.techspot.com/news/78431-human-player-finally-beat-deepmind-alphastar-ai-starcraft.html
I think this happened most likely because just like they did in chess, the AI was trained to prioritize victory over a draw. MaNA took advantage on this and focused on defending and thus getting economical advantage
Re: (Score:3)
I think this happened most likely because just like they did in chess, the AI was trained to prioritize victory over a draw.
It happened because the AI was really, really, good at micro. That was entirely it. The human had to choose strategies that would work even with (relatively) poor micro.
Re: (Score:3)
As the automaton has zero intelligence and absolutely no understanding regarding what it does, the humans would always win after a while.
I've seen enough scifi movies to know that bots are invariably controlled by some central computer, so humans can always win by sneaking in to the main base and blowing up the controller.
Re: (Score:2)
When this type
Is it AI really? (Score:2)
Re: (Score:3)
RTFS
SRSLY, RTFS
The only thing you asked that isn't literally in the summary is whether they pointed a camera at a monitor. I'm guessing they actually used a GPU to render the scene, then handed the frame off to the computer. TFS outright states that humans had better visual abilities (regardless of how the image data was getting into deepmind) but it retained superior reaction time and accuracy.
Your remaining question is answered in one of the links in the summary, but I'm not telling you which one since yo
Re: (Score:2)
"Humans' visual abilities made them better snipers. But at close range, [DeepMind's team FTW] excelled in combat, in part because its reaction time was half that of a human's, and in part because its accuracy was 80 percent compared to the humans' 50 percent."
How the FUCK does that amount to 'ability was adjusted to be equal to a human"? The way I read it, they handicapped it in certain ways and it was still faster than others.
Re: (Score:2)
They handicapped it to add human-level button pushing reaction speeds, but DeepMind is still faster at deciding which buttons to push. Since they're talking about the humans having superior visual abilities, it's clear that DeepMind is having to analyze the scene visually. I didn't find either of these determinations to be particularly complicated.
Re: (Score:2)
Re: (Score:2)
You do know you can look all this stuff up, right? And there's a link to the study in the summary? Oh wait, this is Slashdot. I must be new here.
Re: (Score:3)
Well, I'm guessing they didn't and simply fed it the game state like they do in most of these pointless AI vs human challenges.
If they'd done something as different as having the AI decode the actual frames they would be shouting that from the rooftops.
Re: (Score:2)
If they'd done something as different as having the AI decode the actual frames they would be shouting that from the rooftops.
They've been doing that for half a decade now. Here was the first story on the topic [slashdot.org]. The inputs were the pixels on the screen, and the evaluation metric was the score.
Re: (Score:2)
- Was artificially slowed down and made less accurate to make sure the winning was not caused by aimbot-skills.
It still all seems very ambiguous. What did they even use as a metric to accurately match "human reaction time"? It would have to be pretty exact.
In other words... (Score:1)
Super AI AimBot? (Score:2)
Yea, we have been doing this with much cheaper equipment for a lot longer than this.
What an amazing waste of hardware! Sign me up for an account please! I have some tards to frag!
Re: (Score:2)
Complete nonsense. The concept of a railgun is far older than the concept of a computer game. You are just displaying extreme ignorance.
Games are designed strangely (Score:5, Interesting)
All modern video games are designed to be 'addictive', not difficult nor fun. The term "grinding" is a key example of this.
But in any case, computers have different capabilities. A game that is difficult for humans might require fine motor control, which a computer could have perfect motor control. Similarly, a game that is fun could be because of interesting plot twists based on assumptions that humans make that a computer would not make.
As such, I would expect any competent computer to win all of them.
It is not be difficult for someone to design a game that humans would win and computers would never win. Right now, we call such games, 'turing test'.
Base it on creativity and generalized, non-specific extremely varied pattern recognition without any instruction that is a pattern recognition (i.e. one level could be won by recognizing the music played, another by seeing which icon is slightly different than another, a third by the timing something, all without instructions that any of this stuff matters).
Such a game would not be solvable by a computer, but humans would excel.
Re: (Score:2)
All modern video games are designed to be 'addictive', not difficult nor fun. The term "grinding" is a key example of this.
All Slashdot posts are designed to convey overt displays of sensationalism. A full 100% of them.
Side note: Play better games, there's plenty of damn fun "modern" games out there. Grinding is a negative word used to describe crap games, definitely not an example of "all modern" games.
The fact we are talking about Quake 3 Arena adds a real "WTF" element to your post.
Re: (Score:1)
> Base it on creativity and generalized, non-specific extremely varied pattern recognition without any instruction that is a pattern recognition (
This is exactly what the AI did in this case. It was offered raw pixels of the game and information whether it won the whole match or not. The AI was not told what is a flag nor that the flag needs to be captured, the AI was not told that some moving elements are opponents and some are team mates. Yet the AI learned it all and beat humans in it. Adding music or
Re: (Score:2)
First of all, you are making assumptions after reading a summary article. I bet that what you said is not true, programming is a lot more complicated than that.
Among other things, you said adding music would not be sufficient to change things. WRONG. You said it was told to analyse pixels. That alone is too much instructions. Given that instructions if the game designer added sound as as essential element then the game could not win because it did not not analyse sound. NO. Stop cheating to help t
Re: (Score:3)
The AI was not told what is a flag nor that the flag needs to be captured, the AI was not told that some moving elements are opponents and some are team mates.
Ironically it was basically told all those things.
That's why I like the Souls series (Score:2)
Those games bring modern trappings to old school values. Dragon Quest XI is good too.
I don't like games that have you picking up and opening capsules or chests that you then have to open for random rewards. I'd rather have to hunt in a level for cleverly hidden valuables and thoughtfully placed items in appropriate locations.
Hey duds... you know that... (Score:2)
... really sounds deep, man. Whoa.
I didn't shoot the deputy.. (Score:2)
...and also not the Sheriff, I just started a miniature chemical explosion in a metal encasing and a metal slug at one side was the weak point, the gases went that way and ejected the slug and unfortunately the Sheriff ...
Re: (Score:1)
I used to be pretty good in Quake 1
Buy you're not very good at reading summaries.
If the bots won primarily because of their accuracy rather than e.g. tactics then I'd say this was a waste of time and has not advanced our knowledge or the state of play at all.
We already know that Entities Using an Aimbot are better than Entities Not Using an Aimbot.
The answer to that was mentioned in the summary, you didn't even need to read the article.
Are you sure you're not a bot?
bullshit article (Score:1)
Re: (Score:2)
To be fair, they did move closer towards it by now actually (attempting) playing on actual Q3 maps instead of tiny square-shaped levels.
It's still highly stylized visuall, though and shooting is limited to 'tagging', whatever that may be. No Q3 weapons.
Seminal? (Score:2)
Do you even know what âoeseminalâ means? This was the last first person shooter I spent a lot of time on, mostly because I was bored of the format with Q3.
Wolf 3D. Seminal. Doom, possibly seminal. Doom II, Quake and Quake 2 came from the same company. Unreal took a lot of our time before Quake 3.
We wired our house in 1995 with an Arcnet network (Ethernet too expensive) to play multiplayer Doom.
Q3 was not seminal.
The Wheel of Time (Score:2)
One of the main characters is a general with untold battles, won and lost, in memory. While he is not invincible, he is extremely proficient at war, and becomes more so each iteration.
Wars of the future will be AI against AI.
Does it work on not-before-seen levels? (Score:2)