| Forum Index > SC2 General |
|
ElMeanYo
United States1032 Posts
On November 01 2019 00:08 thewhiskey wrote: Just a little input from a data scientist (studying and working with machine learning) regarding the whole AI debate (by the way, congratulations to the team for reaching this milestone). Very interesting, thanks for your take on this. | ||
|
Quatari
Australia48 Posts
On October 31 2019 08:15 Arrivest wrote: Hi MyLovelyLurker, I'd like to know: 1. When selecting multiple units, human players need to drag a box, which limits how they can micro, for instance in things like splitting marines against banelings. To be fair, could you make AlphaStar drag boxes too in the future? 2. When a dropship shows on minimap, there's a ~1/4 chance a best human player won't notice it. Is this implemented in AlphaStar? 3. When a cloaked unit moves across the screen, there's a ~2/3 chance a best human player won't notice it. Is this implemented in AlphaStar? 4. When clicking on something small, there's a ~5% chance a best human player will misclick, I guess this is already implemented? 5. Will you emulate a virtual mouse in the future, so that we can watch AlphaStar's first person view in the ultimate AlphaStar vs Serral match, which is bound to happen sooner or later? Thanks! First off, I should say that I am not affiliated with DeepMind - the following is just my (unofficial) understanding of how the API that AlphaStar used works. If you want to know for sure, you could examine the API source code at https://github.com/Blizzard/s2client-api. The AlphaStar version on the Battle.net ladder used the RAW data interface (not the feature layers / rendered interfaces). 1. I could be wrong, but afaik, for each action that AlphaStar issues, AlphaStar makes one selection, but I am unclear whether a "selection" means only one of its units, or a set of its units (which could potentially be spread across its camera view, which might not be possible to be selected via just a single box-select without also deselecting some units that it doesn't want to command). I don't know whether DeepMind limit selections to the box-select method. Perhaps someone can answer this? 2. My understanding is that AlphaStar's perception of the visible portions of the map is complete (nothing overlooked), and virtual mouse click precision is perfectly precise according to its intentions (no chance to misclick). Whether you consider this to be a problem comes down to whether you think AIs like AlphaStar should be restricted to the same flaws as human eyes/brains/limbs/control in the real world. DeepMind seem to have decided that it is ok for AlphaStar to be perfectly precise according to the I/O from/to the display/keyboard/mouse (no chance to overlook / misclick). 3. My understanding is that "chance to not notice" is not implemented, although AlphaStar only gets very limited info about cloaked units (e.g. perhaps just its position, not attributes such as its HP. Perhaps also its size). I haven't checked whether it knows its size or unit type (note that a human might be able to deduce whether it is a flying unit or what unit type it is from its shimmer, e.g. on a single frame if they are perceptive, or across multiple frames by changing the camera view in various ways). 4. Afaik, AlphaStar never misclicks. See comments for 2. 5. Just from what I read (I have not watched the replays published by DeepMind), AlphaStar's camera view changes should be visible when watching the replays from AlphaStar's perspective. I don't think the exact mouse cursor positions would be apparent from the replays though - just the camera view. | ||
|
Quatari
Australia48 Posts
On October 31 2019 16:34 heqat wrote: Don't think it is based on game state. From what I understood they have a simplified rendering of the view (like one image for the building, one for the enemies, etc. but it is based on pixels: There are several interfaces available via the API - Raw interface (i.e. structured unit attribute data), feature layers interface (several layers like you showed), and the rendered interface (simply the raw pixel data). AlphaStar just used the Raw interface. The version of AlphaStar that was used on Battle.net was restricted to only being able to access unit data for units that were within the current camera view (not units elsewhere on the map). | ||
|
ThxSub~
17 Posts
On October 31 2019 08:16 tigon_ridge wrote: Why are you people so excited? This organization, which is a child of your big brother google, is engineering the alpha phase of your replacement. How do you people have such little foresight? Skynet isn't just fiction—it's also prophecy. Humans will eventually be outperformed at anything by machines, but you don't need to panic because of that. Because keep in my mind: There are billions of not so intelligent humans on this planet. Also many of them are bad athletes, not very tall or even too tall in some cases. Why is this the reality when in the early 20th century some people had lots of ideas about creating the superior man? Ideas about selectively breeding desirable populations? As mankind we can just say: screw that, this isn't who we are. We can essentially just keep robots as our servant race, instead of giving them out to the free market and letting them compete with us humans, who have to sleep several hours a day and who also want to be more than just work drones anyways. Another anology would be chemical weapons. Very effective (and even cheap I think) but outlawed. Even when the business is about killing each other we can implement certain rules that define our race as something other than just apes who only exist to make everything better and more effective etc. without any regard to sense and purpose. So I worried about A.I. for quite some time too, but there are so many people who are already doing that now. And the more progress that we see, the more people will realize that creating a better species than us is not in our interest. And then we will once again stop doing certain things, just like we stopped doing eugenic programs or using chemical weapons. | ||
|
KT_Elwood
Germany1202 Posts
On October 31 2019 06:59 loppy2345 wrote: Deepmind must be burning through cash on this Starcraft project. Their losses were $570 mil in 2018, and $370 mil in 2017. I feel they could well run out of money very soon, unless they manage to find a commercial use for this type of research. Well if you got to pay a recent graduate in computerscience or engineering you have to spent around 120.000€ in Germany (Thats about 50-60k for the employee and 60k for the employers part of social security cost + cost of hiring and onboarding). In other countries more or less parts of the costs go to the employee directly. Deepmind has 700 Employees, and I would think that most of these would be engineers and computerscientists. So basicly you'd have to spend 84 Million a year to pay a basic salary for 700 people workforce, as Deepmind is part of Google, I'd Imagine they choose professionals and I think they need offices and Computers and stuff, so bam: To be TopDog in AI, you got to spent money | ||
|
BronzeKnee
United States5226 Posts
I watched the longest TvZ game and felt Alphastar seriously underperformed. The micro was excellent, the macro on point, but the predictive ability was terrible. The AI was lured into predictable traps several times. If a progamer knew they were playing Alphastar, they'd never lose. My analysis has proved my earlier hypothesis correct. In a game of limited information, there are too many possibilities and variables to consider for an AI, especially one that lacks Starsense. A human player would have noticed their opponent was setting traps and not been lured more than once. A human player would have been able to predict the tech switches and harrassment much better. But the has AI calculated that chasing down a group of Zerglings or leaving an expansion undefended works most of the time according to its algorithms, so it did so, over and over. It was outmaneuvered time and again in that replay, seeming so amateurish in terms of strategy as it was chewed to pieces. Near perfect micro and mechanics couldn't overcome the Platinum level decision making and I think the human player on the other side probably wasn't impressed. Pretty sure I could cheese the AI to death easily. Assuming equal input limits, meaning Alphastar has to use a mouse, keyboard, and monitor to interact or players can control the game with their mind (which is where we'd see the real beatdown of the AI), Alphastar won't conquer this game. | ||
|
Quatari
Australia48 Posts
On October 31 2019 23:07 heqat wrote: I read this interesing comments on ArsTechnica, I think it resumes pretty well some of the limitation of the experience: "The clicks are still perfectly precise all the time (x,y coordinates). This could be fixed by making the system emulate a mouse and keyboard, with another layer fuzzing the inputs, especially if it increased error as "inputs per second" go up. As a human, if I click 75 times per second, my clicks will not be accurately placed. AlphaStar's still are. No misclicks ever. They apparently did add some input lag on the time domain but it's not clear if this is static, or dynamic in a way that would mimic humans getting worse as we ramp up the intensity. Everyone parroting that they "are using the same APM as humans!" is just, well, functionally wrong. A human gets APMs in the 500-1000 APM range by spamming nearly-useless clicks all over the place. There's another algorithmically estimated benchmark called "EPM" for "Effective Actions Per Minute", and AlphaStar's EPM is still disgustingly high vs humans. The take-home from this is that when AlphaStar beats humans, it still does so using techniques that no human will ever be able to execute. It's not finding new cool strategies that we can learn from. It's not teaching humans how to play better StarCraft, like AlphaZero/LeelaChess Zero teach humans how to play chess better. It's more like, a really really hard-to-detect aimbot that plays "almost" human so if you're playing it maybe you don't notice but when you watch the replay to go "how did he do that?" you go "oh he was using an aimbot, nothing for me to learn here". Also, I've downloaded and watched/analyzed a lot of its replay files. AlphaStar still learned SO MUCH of its play style by simply copying other humans that it was fed. It's still definitely not learning "from the ground up" like AlphaZero learned Chess. My biggest takeaway is that I didn't realize just how many orders of magnitude more difficult StarCraft is, than Chess. I thought it was "somewhat" more difficult, but it's actually an "incredibly" harder game to play, evidently. Right now, it appears that AI simply cannot teach itself StarCraft from scratch, when limited to human abilities. It needs to learn from humans first. As a result, it's the AI's "final form" includes many of the idiosyncrasies common in human play, who sometimes do certain things every game "just for fun", out of superstition, or just weird ticks of neurological habit/reinforcement. A good example of this is AlphaStar destroying a small environmental object ("unbuildable plates") just outside a base. Humans do this because if it's still there during a big attack, they might accidentally click it instead of an enemy unit, so clearing it away early occasionally helps later. AlphaStar can't accidentally click it, so the only reason the AI destroys these plates is because it was fed a million games of top level humans to mimic, and doing it never makes alphastar lose so it hasn't ever stopped doing it. The applicability of this to self-driving cars is muddy. Self-driving cars don't need to limit themselves to human input/output capabilities. In that case, they're free to react as quickly, often, and accurately as they can. As a result, they may find that there are "non-human" strategies for driving a car which work wonderfully well, just as the first iteration of AlphaStar found when it used super-human clicking to bully some of the best humans on the digital battlefield. Lastly, AlphaStar appears to be able to play many different "styles" of StarCraft, but only because it's actually a collection of many different AI's (agents) which grew in different directions. Each individual agent can apparently only play one style (albeit very well). We haven't yet seen AlphaStar adjust its grand strategy mid-game as it's getting beaten. Humans who have mastered multiple strategies can go "oh wow this guy is countering my mass ground unit perfectly, so I'll fake my next attack wave but actually be secretly changing my economy to air units in the background". We haven't yet seen AlphaStar do anything like that. It's possible that with "nearly-perfect' play, that type of game play would be sub-optimal, as it relies on essentially a psychological trick -- hoping your opponent doesn't realize you're making yourself temporarily vulnerable during the strategy transition. But AlphaStar isn't even playing more optimally than humans yet, and humans still do well by changing strategies mid-game at the level that AlphaStar is currently playing, so at the very least, it's concerning. It seems likely that because of the "multiple highly specialized agent" architecture, that AlphaStar probably simply CANNOT switch strategies in the middle of a game. It has no mechanism to select a different agent and hand off control to it, let alone merge the playstyles of two agents to play a hybrid style." Just Re. APM, for clarity, I think it's worth mentioning the limitation that DeepMind said they used in the version of AlphaStar that played on the Battle.net ladder - quoting from the blog post: > Agents were capped at a max of 22 agent actions per 5 seconds, where one agent action corresponds to a selection, an ability and a target unit or point, which counts as up to 3 actions towards the in-game APM counter. Moving the camera also counts as an agent action, despite not being counted towards APM. ... and quoting from the paper: > APM Limits: > Agents are limited to executing at most 22 non-duplicate actions per five second window. Converting between actions and the APM measured by the game is non-trivial, and agent actions are hard to compare with human actions (computers can precisely execute different actions from step to step). As I said in an earlier post, I am unclear whether a "selection" means only one of its units, or a set of its units (which could potentially be spread across its camera view, which might not be possible to be selected via just a single box-select without also deselecting some units that it doesn't want to command). I don't know whether DeepMind limit selections to the box-select method. Perhaps someone can answer this? FYI, for completeness, in the paper, TLO's official statement regarding fairness was: > Professional Player Statement > The following quote describes our interface and limitations from StarCraft II professional player Dario “TLO” Wünsch (who is part of the team and an author of this paper). > The limitations that have been put in place for AlphaStar now mean that it feels very different from the initial show match in January. While AlphaStar has excellent and precise control it doesn't feel superhuman - certainly not on a level that a human couldn't theoretically achieve. It is better in some aspects than humans and then also worse in others, but of course there are going to be unavoidable differences between AlphaStar and human players. > I've had the pleasure of providing consultation to the AlphaStar team to help ensure that DeepMind's system does not have any unfair advantages over human players. Overall, it feels very fair, like it is playing a `real' game of StarCraft and doesn't completely throw the balance off by having unrealistic capabilities. Now that it has limited camera view, when I multi-task it doesn't always catch everything at the same time, so that aspect also feels very fair and more human-like. | ||
|
opisska
Poland8852 Posts
| ||
|
CoupdeBoule
73 Posts
In what way has the AI changed our understanding of SC strategically or tactically? Thats right - in no way whatsoever. | ||
|
General_Winter
United States719 Posts
On October 31 2019 06:41 Big-t wrote: Nice, can Alpha Star play at WCS tournaments?^^ And if it does would it count as an NA player? The company is US based, and frankly we need all the help we can get in WCS. | ||
|
Musicus
Germany23580 Posts
On November 01 2019 02:15 opisska wrote: Where do you people get the replays to watch from? https://deepmind.com/research/open-source/alphastar-resources Link to replays is at the top in yellow. | ||
|
opisska
Poland8852 Posts
On November 01 2019 04:21 Musicus wrote: https://deepmind.com/research/open-source/alphastar-resources Link to replays is at the top in yellow. Oh cool thanks! | ||
|
Musicus
Germany23580 Posts
| ||
|
Amui
Canada10567 Posts
On November 01 2019 02:23 CoupdeBoule wrote: This is all well and good and interesting but Starcraft is so ill-suited for AI projekts because it’s not a turn-based game. In what way has the AI changed our understanding of SC strategically or tactically? Thats right - in no way whatsoever. The better way to think about it is how has developing an AI for SCII advanced our understanding of AI. You're thinking about it the wrong way around, because the researchers are not trying to redefine how to play the game. | ||
|
CoupdeBoule
73 Posts
On November 01 2019 06:05 Amui wrote: The better way to think about it is how has developing an AI for SCII advanced our understanding of AI. You're thinking about it the wrong way around, because the researchers are not trying to redefine how to play the game. Yeah I get that - that is their goal. Guess Im just perplexed why even bother mentioning this project in the same sentence as SC. This is AI stuff - literally has had no impact on SC. | ||
|
WombaT
Northern Ireland27478 Posts
On November 01 2019 06:38 CoupdeBoule wrote: Yeah I get that - that is their goal. Guess Im just perplexed why even bother mentioning this project in the same sentence as SC. This is AI stuff - literally has had no impact on SC. People are interested in the process? Threads are rather active when AlphaStar is discussed, YouTube vids of games vs it have tons of views. | ||
|
Kitai
United States886 Posts
I would also like to take this opportunity to welcome our new AI overlords. Maybe I'm naive, but I see AI as beneficial in so many ways and not as some sort of existential threat waiting to annihilate us all. I also say this as a human being. Do not fear the robots, fellow humans. | ||
|
Jan1997
Norway671 Posts
| ||
|
WombaT
Northern Ireland27478 Posts
On November 01 2019 07:01 Jan1997 wrote: I suppose this is proof that macro is basically everything in this game. Making units is a good skill to have. I’d say one of its biggest strengths that isn’t mentioned quite as much is its choice of when or when not to engage. Seems to know if its army beats what it’s facing or not very accurately. | ||
|
cizzy819
4 Posts
On October 31 2019 06:15 MockHamill wrote: I have a high IQ but can not escape diamond. AlphaStar has less intelligence than a dog but is GM. Conclusion: I must balance whine harder. Blizzard is clearly at fault. Being an annoying hamster has to factor in to these details. | ||
| ||
StarCraft 2 StarCraft: Brood War Dota 2 Counter-Strike Super Smash Bros Heroes of the Storm Other Games tarik_tv5803 Grubby2845 FrodaN1069 Fuzer  194C9.Mang0191 PiGStarcraft170 ArmadaUGS151 mouzStarbuck140 edward80 NeuroSwarm68 KnowMe47 Railgan29 JuggernautJason17 ZombieGrub15 Chillindude10 |
|
IntoTheTV X SOOP
Reynor vs Solar
ByuN vs SHIN
Clem vs MaxPax
herO vs Rogue
IPSL
WorsT vs nOmaD
Dewalt vs andersOn
AI Arena Tournament
IntoTheTV X SOOP
IPSL
Azhi_Dahaki vs ggnore13
Double vs WIZARD
Replay Cast
Replay Cast
Sparkling Tuna Cup
WardiTV Invitational
YoungYakov vs Lambo
SHIN vs TBD
Percival vs Shameless
Solar vs TBD
PiGosaur Cup
[ Show More ] |
|
|
EPT
