Slashdot Mirror
OpenAI Is Beating Humans At 'Dota 2' Because It's Basically Cheating (vice.com)
Motherboard's Matthew Gault provides another possibility for how OpenAI's bots managed to beat professional human players in two consecutive games of Data 2. Gault argues that "it was only possible thanks to significant guardrails and an inhuman advantage" -- not necessarily because the AI was more clever than the humans. From the report: The OpenAI Five bots consisted of algorithms known as neural networks, which loosely mimic the brain and "learn" to complete tasks after a process of training and feedback. The research company put its Dota 2-playing AI through 180 days worth of virtual training to prepare it for the match, and it showed. However, the bots had to play within some highly specific limitations. Dota 2 is a complicated game with more than 100 heroes. Some of them use quirky and game-changing abilities. For this exhibition, the hero pool was limited to just 18. That's an incredible handicap because so much of Dota 2 involves a team picking the proper group composition and reacting to what its opponents pick. Reducing the number of champions from more than 100 to 18 made things much simpler for the AI.
The OpenAI Five bots also played Dota 2 by reading the game's information directly from its application programming interface (API), which allows other programs to easily interface with Dota 2. This gives the AI instant knowledge about the game, whereas human players have to visually interpret a screen. If a human was able to do this in a competitive match against other humans, we'd probably call it cheating. Even with this AI advantage, Walsh and his team beat the bots in the third game, when the match organizers turned hero selection over to the crowd, which gave the AI a weak hero composition. Walsh thinks he and his team could eventually beat the AI in a fair right, even given the limited hero pool and other restrictions.
The OpenAI Five bots also played Dota 2 by reading the game's information directly from its application programming interface (API), which allows other programs to easily interface with Dota 2. This gives the AI instant knowledge about the game, whereas human players have to visually interpret a screen. If a human was able to do this in a competitive match against other humans, we'd probably call it cheating. Even with this AI advantage, Walsh and his team beat the bots in the third game, when the match organizers turned hero selection over to the crowd, which gave the AI a weak hero composition. Walsh thinks he and his team could eventually beat the AI in a fair right, even given the limited hero pool and other restrictions.
what many of us pointed out in the comments when the story originally ran here on /.
I find applications of machine assistance more promising. I'm still waiting for a way to program the original starcraft so that the AI can manage the tedious resource management while a human player can work on strategy. Even something like self driving cars make more sense as an assist. For someone who is elderly do they really care if the car is being driven by a computer or by a human sitting in an office somewhere. If the human sitting in the office gets the advantage of an AI highlighting the road for them and the elderly only has to pay the human for the 10 minutes to and from the appt. The reason chauffeurs are so expensive is because you have to pay them to sit. If you have AI drive the interstates and a human drive the final 10 minutes on each end, it would be very economical.
Indeed. This is an early iteration of the system. Eventually with enough effort, visual recognition will be at sufficient level, there will be a mechanical hand to press keyboard buttons and manipulate the mouse, and AI will have training to do go with all the heroes.
Saying that "but we can win it" forgoes a critical caveat. "For now".
Cheating means that true Artificial Intelligence that mimics humans is here!
The shepherds did so well protecting the flock that the sheep no longer believed that wolves existed.
Dota is a complicated game but its a 5v5 and a hive mind will always be more organized than five individuals.
More importantly there are items to Hex, Silence, Disarm and otherwise disable other characters and a fight can be decided on who gets disabled first.
Human beings have mental reaction time and physical processing time of moving a mouse to the right coordinates.
Computers have none of that and they will always click on you before you can click on them. That doesn't make them smart.
Really? The best they could come up with is comparing the bot to a programmable mouse, illegal in human play? This standard basically disqualifies any bot implementation being fair in their definition just on the basis of being a program. What do they want? The bot to send commands to a human operator which plays the game for it?
Their slightly less ridiculous claim is saying the bot is unfair because it interfaces with the game not through mouse and keyboard and the dota renderer but through valve's bot api. Just..fine, if you dont care about how good a bot can be under the dota developer's own definition, then ignore the results.
I just dont understand the mindset here, is this resistance to change? People dont want AI to be good, so they make up narratives to explain away recent successes? It really feels like the same mindset as conspiracy theorists.
Do you know the difference between DOTA and DATA?
Neither does BeauHD.
...Given enough computing power, a self learning NN will beat humans at any task.
When will AI be good enough to give us shit articles, I wonder?
His ignorance covered the whole earth like a blanket, and there was hardly a hole in it anywhere. - Mark Twain
No mention of the fact the courier visibility to the enemy team had to be turned off as well because the AI was so bad at using them?
...routinely, for probably close to 3 years now. Still don't have a fucking clue what its actual name is.
I believe it is "Day Of The Asshole".
But I can't google either.
His ignorance covered the whole earth like a blanket, and there was hardly a hole in it anywhere. - Mark Twain
If you'd been paying any attention at all, you'd know that the AI's latency was set to 200ms, which is larger than the average human's.
Same as these guys -- their own logic is self-contradictory. Either DOTA is a game mainly about reaction time, in which case the 18-player limit will have almost no effect; or DOTA is a game mainly about strategy and how to use characters together, in which case the direct interface will have little effect. Given the fact that poorly-chosen characters caused the computer to lose decisively, I think the first one is much more likely.
The OpenAI team have stated over and over why they use the direct interface rather than scanning the pixels: Because they know how to get an AI to scan pixels, but they don't know how to get an AI to do strategy. So they're focusing their training time on strategy rather than wasting GPU time scanning pixels.
TCP: Why the Internet is full of SYN.
In less time than it took you to type that pointless moronic comment you could have googled it and answered the question
for instance, the performance of AI on simple video games is due, in part, to the fact that it simply has less latency in scanning the screen, pushing that information into a meat cortex, and pushing its meat appendage against an ergonomic device. this doesn't detract, imho, from the accomplishment (however much you may or may not think of it as significant in the first place).
as for cheating: well, i think people took a tech demo maybe a bit too seriously... it's still impressive imho, even if it was playing a "special edition" of dota2. i've seen much, much more incomplete proofs-of-concept get obscene funding.
"They were pure niggers." – Noam Chomsky
yeah, i read that: "Machine Control of Semi-Autonomous Combatants in Simulated Areas of Conflict."
DoD money is great, isn't it?
"They were pure niggers." – Noam Chomsky
Who is Matthew Gault?
Corruption is convincing someone that the selfless ideal is the same as their selfish ideal.
It's provably false: these neural networks are pattern matchers, not Turing machines. This means that you can throw as much computing power as you want, they won't have general intelligence.
"First they came for the slanderers and i said nothing."
An airgap should be required. The machine only sees the game with a camera/pair-of-cameras. The machine only makes moves osing electromechanical servos placed over a controller/keyboard.
In less time than it took you to post your scold, you could have defined the DOTA acronym.
And they are already doing just that with classic Atari games like Breakout. The only input the AI got was the score and the raw screen pixels. It had absolutely zero information about the fact that there was a ball that needed to be bounced up by a paddle. It just had to figure out that the white pixel on one frame and the white pixel in a different position on the next frame were somehow related, and something bad happened when it reached the bottom. After a while it was breaking bricks on one side to get the ball on top of the bricks, surprising even the researchers who had not expected this kind of behavior.
You still need to equalize on input end as well. Humans have to go through man-machine interface that is keyboard, mouse or controller. AI is typically directly plugged in.
So you'd need some kind of a mechanical manipulator for a keyboard/mouse/controller to fully equalize both input and output. In your scenario, it's just output that is being equalized. Your experiment only goes half way of the necessary road for what could be considered an equal playing field.
Why did you crapflood this irrelevance? It wasn't 'sly sarcasm' because those aren't direct factors in game play like what I posted about.
If the neural network had access to a clock, I would expect it would eventually find this strategy. Time between hits would directly correlate with lowered risk of failing the game
Even those who arrange and design shrubberies are under considerable economic stress at this period in history.
A lot of the brain has many pattern matchers, so it is likely to be a prerequisite for GI, or may simply be an emergent property of a sufficient number of them, wired appropriately. We don't yet know, as actually building something with the complexity of the brain with feasible power budget is still difficult, and within a reasonable energy also difficult, and would result in slow thought. In a decade it may change and be at least possible. Another decade and GI may be demonstrated.
I watched all the matches. The AI 'reading' the API didn't stop them making fucking 'dumb' decisions no human player would ever make. Buying salves(consumable healing item) constantly, wasting smokes randomly, double ward glitching. The only questionable play they made was the near instant hex on a blinking earthshaker about to land his combo. But top level players know what the fuck earthshaker does.
They are prepared for him to show up like that. Its called fucking learning, and its exactly what this AI did. You can talk about unfair reaction speed, but experience means far more, holding your disable cooldown for the right target.
Making the correct decision is not fucking cheating. They acted without hesitation, and they crushed m-dog, f-dog and every other dog that dared taunt them with 'baited kid'. Up until they let 4chans IRC aka twitch chat draft their 3rd game.
It's definitely not am emergent property of building a lot of them, and we do know because it is provable. I encourage you to read a book about AI and not read stupid articles. You are right that the brain has a lot of pattern matchers.
"First they came for the slanderers and i said nothing."
It's definitely not am emergent property of building a lot of them, and we do know because it is provable.
I think you need to read what I wrote: "wired appropriately", which refers to the potential interfaces between them and the elements which are not pattern matchers. Also note the word of the word may, but to be fair, I should have emphasised the 'glue' elements again in that sentence, so I was a bit sloppy and my words don't fully reflect what I was trying to say, so I can understand some confusion.
I encourage you to read a book about AI and not read stupid articles. You are right that the brain has a lot of pattern matchers.
I've read several, and about the workings of the brain, as I was, for many years an AI researcher, and also worked with neuroscientists. One of my good friends is an AI researcher with a PhD in neuroscience. Maybe I will bounce a few ideas off him.
I've read several, and about the workings of the brain, as I was, for many years an AI researcher, and also worked with neuroscientists. One of my good friends is an AI researcher with a PhD in neuroscience. Maybe I will bounce a few ideas off him.
Great, then you should understand that these neural networks are not Turing complete. You should also understand what it takes to make them Turing complete, and the drawbacks of doing so, and why those types of networks have not been as successful.
"First they came for the slanderers and i said nothing."
Great, then you should understand that these neural networks are not Turing complete.
Indeed, and I very nearly made that point in my previous post. I also noted that I wasn't talking about a single artificial neural network, but an assembly of them. That research has not yet does not mean that GI could not be an emergent property of an assembly of them. It may later be show that it is not possible, but the jury is pretty solidly out on them at present. The current neural networks do some limited functions extremely well, but there is a lot of work to do to emulate the complexity of a brain, which is our current model for intelligence, given the lack of a watertight definition on intelligence.
However, there's a worry that there is a god of the gaps argument emerging, as various tasks that were believed to require GI have been shown to be achievable without this. This doesn't mean that such tasks can't be achieved by using a more general intelligence, as obviously they can, but at one point it was assumed that they required it. Thus I would not be tempted to make absolutist statements about whether GI could or could not emerge from collections of pattern matchers, as if might be a poor bet.
However, there's a worry that there is a god of the gaps argument emerging, as various tasks that were believed to require GI have been shown to be achievable without this.
This has happened when it was shown that the task is simpler than previously thought. For example, with the Turing test........Turing never expected it to be so easy to trick the average human being. Likewise, with computer chess, all you need is a really powerful tree searching computer. No intelligence needed.
"First they came for the slanderers and i said nothing."
I did a spit take when I read "Dota 2 is a complicated game". Starcraft is a complex game. Sure there are 100 heroes in Dota, but only 10 of them can be on the field at once (I think, I don't play it so please correct me if I'm wrong). In Starcraft you can have more than 400 units on the field at once, all doing different directed actions.