|
On January 25 2019 06:44 Waxangel wrote:Show nested quote +On January 25 2019 06:43 Poopi wrote:
I think the title is kinda misleading.
Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.
Perhaps a more compelling nitpick for you then would be how TLO was offracing. He is not pro level with his Protoss. Really, it went 5-1 against pros, and I don't even think parent's argument is insignificant. Also, Mana seemed to find a major flaw in the final game, warp prism harass. I'd imagine in it's current state if you did this strategy 100 times, the bot would lose all of them. Title misleading to seem more impressive than it truly is. Humanity will not kneel to machine.
|
I really don't understand why TL is playing into this PR stunt which only works on morons.
User was warned for this post
|
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
In the near future people will demand “the computer use the same amount of bio-mass as a human brain for its calculations otherwhise it’s unfair”...
Even the “unfair” super-human blink micro was self-learned by the AI. That alone is pretty impressive, given that 1-2 years ago Neural nets could barely win micro fights against the built-in AI.
|
On January 25 2019 11:02 imp42 wrote:
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
In the near future people will demand “the computer use the same amount of bio-mass as a human brain for its calculations otherwhise it’s unfair”...
Even the “unfair” super-human blink micro was self-learned by the AI. That alone is pretty impressive, given that 1-2 years ago Neural nets could barely win micro fights against the built-in AI.
The point here instead is indeed to possibly have an AI beating a human while not relying on superhuman skills; its capability of taking autonomous decisions and developing original and appropriate strats is what it counts, not perfectly microing blink stalkers to the point humans can't keep up: every "stupid" ordinary AI can do this already.
Thus said, the AI not being coded to do so but getting to that by itself is truly impressive; it just isn't what Deepmind project is developed for and expected to do(I guess).
|
United States33499 Posts
On January 25 2019 10:47 CicadaSC wrote:Show nested quote +On January 25 2019 06:44 Waxangel wrote:On January 25 2019 06:43 Poopi wrote:
I think the title is kinda misleading.
Raw interface AI goes 10-0, camera interface AI goes 0-1, would be more appropriate This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride. Humanity will not kneel to machine.
If you said that unironically I guess we've truly arrived in a new age
|
On January 25 2019 09:26 swissman777 wrote:
Also does anyone know how to open the replays? I can't seem to be able to open them.
me neither
|
On January 25 2019 11:02 imp42 wrote:
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
In the near future people will demand “the computer use the same amount of bio-mass as a human brain for its calculations otherwhise it’s unfair”...
Even the “unfair” super-human blink micro was self-learned by the AI. That alone is pretty impressive, given that 1-2 years ago Neural nets could barely win micro fights against the built-in AI.
This is indeed extremely... let's say weird. But is more disturbing is this stuff being pushed in 2019. No one thinks that such a computer can lose to a human without limitations. In the end, it becomes a pointless exercise of giving random limitations to the computer while still having it succeeding. That is why I called this a moronic PR stunt. It accomplishes nothing, the outcome is not unexpected at all (I hope so, at least).
This whole thing would be interesting if somehow you could put real human limitations on the AI so that we could get something interesting out of its strategic decisions. If we wanted to see meme micro we could just get one of those years old videos where a computer would dodge tank shots with a huge group of zerglings but isolating the single zergling that was being targeted.
|
Norway839 Posts
100% confident that AI will drastically outperform humans no matter what, even if you impose further mechanical restrictions such as mouse boxing delay limitations, minimap mouse movement lag (to name a few components resulting in more fair army movement), certain alterations to micro ms delays, etc., Basically even if you model the AI handicaps as close to human level as you can, strategically it WILL beat humans (imo). If today's demonstration and what we've seen earlier with AlphaGo doesn't make people respect the potential of AI then I don't know what will really.
I'm glad I'm not one of those believing that humans can make better decisions than machines in games that are purely down to math and attention/movement. AI will eventually be better at pretty much everything that a human can do. It doesn't matter if you bring in your favorite top players either, Korean or not. Maybe in the short term. But betting against the machines long-term (5+ years) for something as 'simple' (maybe not "simple" in our current year, but soon enough) - as a game of SC2, just seems unwise to me.
The real issue will be down to modeling the mechanical limitations of humans and drawing the line at certain averages, trying out different sliders for different properties, and tightening when the AI makes moves that should require additional enforced millisecond "movement cost". But I'm pretty sure that even if you have a really tight model that moves as sluggishly as a human - especially in multi-location skirmishes - the AI's ability to calculate the board state and "what to do next" will absolutely destroy that of a human. Eventually.
Great work by the deepmind team and professional showing by everyone involved in the stream today. It's clear that the handicaps have to be modeled pretty tightly for the AI to be considered fair to us humans, but this was very impressive overall. 
|
On January 25 2019 11:02 imp42 wrote:
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
You've got it exactly backwards. The people critical of this result are interested in the computer playing by the same set of rules, right now it's playing with extra rules that benefit it, which are not available to humans. Allowing it to see without the restrictions of a camera is a special rule that benefits the computer. Allowing it to have direct API access is a special rule that favors the computer.
Starcraft is a game designed and balanced around physical limitations of the input & output devices used to play it. If you're not using those devices, or at least realistically observing their limitations, you're not playing starcraft.
If you're impressed with the results, then hey, great. But it's not fair to say critics are the one demanding special privileges here.
|
AI feels like such a terrible, terrible idea for humanity. I will never address a robot by a human name (renamed my "Alexa" to Echo) and I will never play VR games.
|
On January 25 2019 11:28 No_Roo wrote:Show nested quote +On January 25 2019 11:02 imp42 wrote:
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
You've got it exactly backwards. The people critical of this result are interested in the computer playing by the same set of rules, right now it's playing with extra rules that benefit it, which not available to humans. Allowing it to see without the restrictions of a camera is a special rule that benefits the computer. Allowing it to have direct API access is a special rule that favors the computer. Starcraft is a game designed and balanced around physical limitations of the input & output devices used to play it. If you're not using those devices, or at least realistically observing their limitations, you're not playing starcraft. If you're impressed with the results, then hey, great. But it's not fair to say critics are the one demanding special privileges here.
At the end of the day, Deepmind is not at all interested in Starcraft. What they are interested in is a controlled environment of imperfect information where real-time decisions must be made from a near-infinite set of choices to achieve a desired outcome.
Starcraft just so happens to fulfill those requirements while also providing a handy PR platform, and whether or not AlphaStar is actually conforming to the exact specifics of some fan expectation of how it "should" play is a tertiary consideration at best.
Deepmind is an engineering company, not a Starcraft company. Priorities follow as such.
|
On January 25 2019 09:46 snakeeyez wrote:
Have they said what kind of neural net it uses? How big is the neural net? They said it also uses reinforcement learning.
About what kind of neural networks, Oriol Vinyals (the first person from DeepMind team) said in the video and also in the blog post that AlphaStar's core structure is based on LSTM (Long-Short Term Memory network, a type of RNN - Recurrent Neural Network), and interestingly from the blog post, the other part is a rare type of structure proposed by Google Brain years ago called Pointer Networks for finding action sequences efficiently. That is even though the reaction time (or what we like to call inference time as mentioned in the video) is about 300 milliseconds on average for an agent, but instead of output one action at a time, it will be able to produce a sequence/batch of actions from one output. Hence it is able to execute APM up to 1500, that is 40 milliseconds per action, which means one AlphaStar agent output probably consist around 10 sequences more or less (think about it like human reflex reactions during micro, instead of every single action is a conscious decision, you make a decision, and the rest is auto execution from muscle memory, which in our human brain is controlled and learned with Cerebellum)
As to how big is one individual agent, from my own experience in evaluating neural network inference time (using one top of the line Nvidia 1080 Ti, and 2080 Ti, and P100), at the scale of millions to tens of millions of parameters (weights in the networks) per network will be around 100 milliseconds per output/inference (without optimization and using Tensorflow). So I imagine one single agent in AlphaStar is probably on similar scale. And as shown in the blog post, the training process not only optimize the parameters but also hyperparameters, hence I suspected the end result size for each agent will be varied, and is make sense to evolve these agents to be efficient in decision time to limit the networks to be compact.
As to how these agents are trained, from what I can gathered it sort of like evolutionary algorithms (a big branch of AI, which is not entirely related to how neural networks are traditionally trained). On this part, the self emulating and keep variations within the agent populations are the keys to generate novel strategies, more so than each network structure. It essentially only need agent that can evolve through many generations where starting populations learned from human records in the virtual league can learn and improved upon themselves, as well as new generations introduces over and over.
|
On January 25 2019 11:20 xuanzue wrote:Show nested quote +On January 25 2019 09:26 swissman777 wrote:
Also does anyone know how to open the replays? I can't seem to be able to open them. me neither
You need to download their custom map and put it in the Maps folder first, as described in their download page
deepmind.com
To load the replays:
Install StarCraft II. It is free to play and runs on Windows and Mac.
Create the StarCraft Maps directory:
Windows: C:\Program Files (x86)\StarCraft II\Maps
Mac: /Applications/StarCraft II/Maps
Download the map.
Move it into the Maps directory.
Download the replays above, and open them as usual.
|
On January 25 2019 11:25 Liquid`Snute wrote:
But I'm pretty sure that even if you have a really tight model that moves as sluggishly as a human - especially in multi-location skirmishes - the AI's ability to calculate the board state and "what to do next" will absolutely destroy that of a human. Eventually.
This is what I wanted to see, but this is not what was displayed. As Xamo noted before in the hands of the AI the stalker is the best unit because of the shire micro potential. No scouting or reactionary play necessary.
I would like to see a strategy and reactionary play which wasn't displayed here. DeepMind, lower the eapm to make it comparable to human and let it win using strategy, not pure micro management. Don't use raw view.
The agent is an achievement nevertheless, but it didn't actually solve SC2, it exploited it unbalance for the APM it has.
|
On January 25 2019 11:48 counting wrote:Show nested quote +On January 25 2019 11:20 xuanzue wrote:On January 25 2019 09:26 swissman777 wrote:
Also does anyone know how to open the replays? I can't seem to be able to open them. me neither You need to download their custom map and put it in the Maps folder first, as described in their download page deepmind.comTo load the replays: Install StarCraft II. It is free to play and runs on Windows and Mac. Create the StarCraft Maps directory: Windows: C:\Program Files (x86)\StarCraft II\Maps Mac: /Applications/StarCraft II/Maps Download the map. Move it into the Maps directory. Download the replays above, and open them as usual.
the map asks for some login that seems fishy
thank you btw, speedreading is bad for me.
|
impressive, but I would like to see all the matchups, esp. no mirror, and different maps
|
Maybe I'm wrong about this but after watching the replays I really think that AlphaStar has discovered a novel way to open in PvP that is essentially superior to what humans are doing currently. The probe production isnt actually a mistake at all. The extra workers make the opponent's harassment pay off way less and can even be used to defend aggression. Once your natural is up, the faster saturation on 2 bases kicks in and you go into the mid game with a supply lead.
I actually think its early game play is extremely intelligent and demonstrated a clear understanding of the strategy.
|
On January 25 2019 11:28 No_Roo wrote:Show nested quote +On January 25 2019 11:02 imp42 wrote:
Does anyone notice the irony of demanding a handicap every time a computer does something better than a human and then arguing that the computer is not better?
You've got it exactly backwards. The people critical of this result are interested in the computer playing by the same set of rules, right now it's playing with extra rules that benefit it, which are not available to humans. Allowing it to see without the restrictions of a camera is a special rule that benefits the computer. Allowing it to have direct API access is a special rule that favors the computer. Starcraft is a game designed and balanced around physical limitations of the input & output devices used to play it. If you're not using those devices, or at least realistically observing their limitations, you're not playing starcraft. If you're impressed with the results, then hey, great. But it's not fair to say critics are the one demanding special privileges here.
Maybe take an additional step back and notice that the game itself is already extremely unfairly biased towards humans (see https://www.teamliquid.net/blogs/518977-towards-a-good-sc-bot-p5-hi-2-2, section 1). From this point of view, we already play with a lot of rules that benefit us humans.
Regarding the zoomed out camera view, I agree. That's clearly an unfair advantage for the AI. However, first results seem to show that having to control the camera just slows down the learning a bit and does not represent a significant obstacle.
Direct API access is a different story. It would be relatively easy to engineer a robotic arm that physically uses a keyboard. It's just not valuable in terms of advancing the field of AI.
|
On January 25 2019 11:52 ivanRT2 wrote:Show nested quote +On January 25 2019 11:25 Liquid`Snute wrote:
But I'm pretty sure that even if you have a really tight model that moves as sluggishly as a human - especially in multi-location skirmishes - the AI's ability to calculate the board state and "what to do next" will absolutely destroy that of a human. Eventually.
This is what I wanted to see, but this is not what was displayed. As Xamo noted before in the hands of the AI the stalker is the best unit because of the shire micro potential. No scouting or reactionary play necessary. I would like to see a strategy and reactionary play which wasn't displayed here. DeepMind, lower the eapm to make it comparable to human and let it win using strategy, not pure micro management. Don't use raw view. The agent is an achievement nevertheless, but it didn't actually solve SC2, it exploited it unbalance for the APM it has.
Just let the 5 most robust agents play against themselves. Assuming their blink micro levels are very similar the exploit cancels out and strategy / reactionary play will be the dominant factor again.
The only thing you will be missing is the arbitrary benchmark under conditions that consider the physiological limits of humans. Not very relevant in the grand scheme of things.
I for one would enjoy watching two AIs battling each other with crazy strategies at 15'000 apm! *
* likely in mirror matches only - I am aware that the balancing is targeted at human level apm.
|
One limitation of sc2, is that you can't zoom out freely, so it's very hard to judge army positions when you are multi-army flanking.
|
|
|
|
|
|
EPT
