Artosis is doing a great job here with the explanations. He has a ton of experience with casting, actually belonging to the best caster duo (Tasteless + Artosis) we have in SC2. I can fully agree with you that the "normal" cast of pro games is very unfriendly for beginners, the game is just too complex and you have to take a lot of things for granted to focus on the things that actually make the difference of winning and loosing (for example the detailed explanation of what a unit does and which army compositions are usually seen, you won't find that in a normal commentary). Too bad there isn't much like this on youtube.

Lowko provides a small trickle of background information about unit abilities, costs, etc. in many casts. I've played a bunch of SC:BW and SC2 on the order of ~5 games and I'm entertained by his casts. (FWIW.)

At this point i've probably followed this game for 3-5 years and I remember this map being one of the first ones I ever played in ranked. I'm unsure if pirates cove was the same map pool along with some other map I don't know the name of. Either way, it's very funny to see how beginner friendly Artosis can be despite having talked about starcraft since 1998 and starcraft 2 since 2010, maybe earlier since the beta and al that stuff. Either way it speaks a lot about him here make it so beginner friendly seemingly for people with little to no experience with the game. I smiled ear to ear hearing him not just explain the production tab but also the adept ability 'shade'. That is so fundamental I almost smiled off my jaw. High respect for helping beginners though. I think I probably was pretty adept at the game by the point this video came out but I think I got into it though videos like these, explaining basic ideas, concepts and information from the overlay(s).

@Rick morty I can't believe how ass-pained Starcraft players still are about League, even down to its casting. You can hate the game for all I care, but calling the casting bad is the most ignorant and ridiculous shit I have ever heard from a starcraft elitist. League definitely has the most sophisticated commentating by this point, the casters being able to predict picks and bans in advance regularly and following map-movement before it happens. They work well with replays and the interface (like drawing on the map) and are knowledgable about players, the history and the meta. Just check out the casters notes by Froskurinn for *one* match: medium.com/@lolFroskurinn/group-c-funplus-phoenix-splyce-jteam-gam-esports-dcf2ef11e903 It really pisses me off, because I like both games. But Starcraft fanboys are such huge bitches with an inferiority complex towards a game, that they can't even get to appreciate any efforts made. Riot has stepped up with the professionalization of their events year after year after year. Their casters are top-notch and have helped to create some of the most memorable memories for me in gaming.

to an extent made a 11 min video 33 mins long

a little bit too slow - I'm a complicate newb to starcraft but this isa gear to slow of intro nbd

league of lesbians is for kids who cant handle the real mechanics of dota. cry more noob

@@preventablesuffering6239 12 year old comments incoming...

Funny how at the same time there's tons of comments from impatient people expecting a regular event cast.

@@genesises Hopefully there is hundreds of those out there already.

I completely agree. This is a great breakdown. I realized how many things we take for granted when I tried to explain the game to a new player recently. It was surprisingly hard to do.

@@salomondeleeuw Knowing what a stalkers abilities are isn't bragging sir.

Yeah, agreed. Thank you for the super detailed breakdown of this elite replay. Forget these nay-sayers!

i can only agree this makes the action soooooooooo boring!

This is very very good for the beginners, which I believe is what he is targeting with this. The researchers who aren't playing the game are very interested.

what if just upload a separate video? one for full gameplay, one for full analysis. so it can also attract other viewers

yep. leaving because of it. Idgaf about your terrible extended narrative.

@@seanruffolo6083 No one cares if you leave.

Too much explanation

Explanation is ok, but pausing is not :(

@@szymonsabat8272 you are right. I guess that's really what I didnt like - the pauses. If he had been able to cram or shortened the explanations without pausing the game it would have been ideal.

wrgg

That's a really interesting point, because it adresses why StarCraft was interesting for DeepMind in the first place: SC is a game of strategy, and strategy is a big part of all the Alpha-projects. Go, chess, and now StarCraft. Like the commentator said; most pros focus on making a lot of those units that are strongest and deal the most damage, but Alpha does not because it focused on the strengths of the other units as well and found something better. In short, it found the best strategy.

@@bard6184 I disagree. AlphaStar does not find the best strategy, it find a strategy with the highest correlation with victory and gets to try it against an endless stream of pro players on ladder that are mostly unaware they are facing the AI. If players are allowed to face the AI repeatedly in that fashion, I suspect literally every strategy in SCII is exploitable (or near it, and any that is not is a clear red flag for game balance), so we would see more AlphaStar losses and DeepMind would be forced to stop pretending they're even better than they really are. Seriously, they've done a good job with AI, but their overhyped claims go beyond their real achievements, not only with AlphaStar, but with AlphaFold as well. I'll leave the Chess and Go assessments to those more familiar with DeepMind's claims vs. achievements in those areas, only speaking on what I've taken the time to look at critically myself.

@@zvxcvxcz I'm not so sure. Alpha gets stronger the more games it gets to play. Leela Chess, which is basically the same AI, the only difference being that it has played a lot more, is far stronger than Alpha ever was. But that is chess and this is SC, and I don't know much about SC. These facts may not apply to SC as it has in chess.

@@bard6184 Ai also has better epm than humans due to them being ai's which is why they can use units such as disruptors a better than pro players.

Play desert strike in arcade mode it’s stress level is lower and it can be a great intro to unit counters if you do ever want to play.

Agreed. It's amazing how people say the restrictions are "human level" when it can target and pull back multiple probes, from a tight ball, while flicking between screens and unit groups and harassing on the other side of the map. There are just so many aspects that won't be on the same level until it's forced to play with a mouse and keyboard.

@@TheBbaughman there is an entire paper on why Alphastar´s "wits" cannot be properly gauged in starcraft 2 due to how much of the game´s outcome is determined by a superior control of the game. Many of the human limitations dont apply to alphastar (atleast yet), so any accomplishment in regards to its neural network, is overshadowed by its superior control of the game. Even on just 180 apm (which is just 3 commands per second), alphastar is gonna have better precision on issuing these 3 commands per second... perfectly hitting the exact pixel where it wants specific units to go. We have known for a long time that robots are better suited for precision work. The automobile industry, the chip industry... they all have been using robots to do precision work in lightning speed (and here physical limitations akin to humans somewhat still apply). In a digital setting, not even these physical limitations apply, thus the robot obviously becomes even more efficient in precision work (which ordering around units basically is). But "wits" is what alphastar wants to demonstrate, yet it cannot really do that without all other degrees of freedoms having been removed. And the most upsetting degree of freedom obviously is lightning fast, mistakeless, precise control. Any "control" mistakes it does, are not due to its hand slipping, or its eyes getting tired, or its fingers being too slow or losing precision due to an adrenaline rush during an engagement, they happen due to alphastar not knowing that they are mistakes. Those mistakes are purposefully executed. Whereas a human can do a mistake and KNOW that it is a mistake, but realize that he simply wasnt able to control it more efficiently for better results.

@@gh0un That's not true though. Alphastar will do misclicks and mistargeting, occasionally will attack its own units for one second. In this case, its that it can see each unit clearly in the ball, but it is still capable of misclicking actions.

@@DaRealDJ Not it does not misclick. If it does that, it does so because it thinks it's useful to do so (even if it isn't). Each click is purposefully executed, there are no control mistakes. It's just a bad decision to attack your own unit, but alphastar doesn't always know that at all stages of the game. It might have figured out that it isn't beneficial to do so in the early game, doesn't mean it knows not to do that in the late game. That's not a case of bad control, it's a case of bad decision making.

@@gh0un I agree with ur comments yet part of the goal of training an AI is to beat human. Being able to achieve this goal itself is something worth noting. Machines are designed to be more precise and mistakeless than human just as tigers are born stronger and faster than human. One thing limiting the power of machines/robots previously is that it can never be programmed perfectly. Now AI provided an answer to this and showed that the programming itself has evolved to a point where even if it's still not perfect, it is enough to bring the overall power of machines over human - which is impressive and never achieved before.

Yes but extremely slow with so many unnecessary pauses. I haven't played in almost 10 years and even understand how the new units work after seeing them in action. I'm sure the guys watching this for the AI can figure out intuitively how these characters work without making this video twice as long.

@@zackj997 Not everyone plays the game, some are just here to watch the AI

Technically alphastar being an ai, means it would have superior multitasking ability and ought to have superior apm

@@ayurge This. I remember in one of AS's early stages, it didn't have a limit on its apm and it was reaching upwards of a couple of thousand apm almost continuously. Granted most of its actions it was performing at the time weren't exactly good or efficient actions because it was still learning the game, but still. I think limiting the apm was one of the first things they had to do if they wanted to obtain any valuable data.

@@Enkarashaddam I am curios whether harder to control unit composition, not as popular among pro players as explained, is due to a superior ability to micro? There has to be a limit of how many different classes of units player can effectively control. Spellcasters or disruptors require a little more APM.

overhuman micro

@@HitoryDensetsu Thats because its not human

@@rowanmonk5652, seriously??

On the other hand, if the wall off is good, it would have taken even less damage

@@michaelmao6180, true) And if he was super good in scout, tactics, strategy he can beat Serral too. But instead we see perfect execution of buildorder, dumb wall and super probe micro.

"fun" he says. Hahaha... Yes, from a certain point of view the cocktail of adrenaline, stress, utter disappointment, and a few sprinkles of victorious joy could be called "fun", but I think you have to be a little crazy to make it much past diamond level. It's not for me, but I do love watching it.

this game is like more demanding than chess or alpha go because you need also a lot of mental speed

@@qwp1026 More a ladder player who already knows what creep and bile are

I dont play SC2 at all but I know what all the abilities are and can deduce the reasoning behind tactics while i watch. Its all pretty straight forward. I agree this would be very educational to someone new but i wish this exact series wasnt the one used as an educational vehicle. There are millions of people like me who purely watch SC2 and dont play. It is fast paced excitement for me so yes there are millions of disappointed people who cannot enjoy this extremely important series.

yeah just like how everyone thought alphastar was an idiot to "over saturate" the bases.

Thats next level

This is actually something I learned from these Alphastar games, if you warp in a hostile unit in the mineral line the zerglings can't a-move the probes because they prioritize hostile units or cannons, so even without the wall as long as you have anything hostile in the mineral line it's actually hard to get an efficient harassment with just zerglings. I still think not making the wall is a mistake though, I think because Alphastar perceives the map through a grid, the grid resolution makes it think it's walled on that map when it actually isn't.

Hold position micro beats the hostile unit trick

Take a look at the AlphaGo documentary. It does similar things there.

Man, we came here to look SC2 vids - pro games, pro casting, all the stuff. Do we really need those kidergarden basics? I would like to listen more about macro and players mistakes rather than "this is medivac - flying unit that can carry other land units, or apparently heal them"

@AmplifiedSilence to play devils advocate I dont see the logic in using the Serral/AS game as the vehicle to educate and explain the basics of the games. Why not a random master or GM ladder match? Furthermore if someone clicked on this video with 0 prior knowledge of SC2 then they wouldnt even know what AS is in the first place. This video is important to the community and seems like an inappropriate educational vehicle. Like it or not the constant and prolonged pauses greatly reduce the entertainment of the video. Its a shame.

This level of complete explanation of everything really does kill a lot of potential enjoyment. If I watch a chess match of the best player in the world versus a computer, I don't want to listen to people explain what the pieces do. I thought it went well with the main alphastar video, but the main appeal of more alphastar work is what it comes up with that top starcraft players don't do, the appeal is not the basics

Give AlphaStar the control over a "terran" army, nothing bad can happen! #we#want#SkyNet.

Are you an effective team?

That was the plot of Ace Combat 7 actually

has john connor actually been born?

Yes, because AI use game API for control instead keyboard and mouse.

Yeah, I wonder if they couldn't find some way to create an input layer between the API and the graphical output that the AI is forced to work with so that it simulates actual mouse selection?

@@Zeuts85 the issue is that mouse group selection goes through the display system (rectangle creates a 3D cone-like collider that is based on the camera and the are of the viewport that has been selected that is intersected with unit colliders); but to let the machine "see" the game they (most likely) provide it with the game state itself, that is more like a super-detailed mini-map with all the data. It is a weird point because the whole point of AI is to make it Superhuman, as it is already in Go and Chess, if it can act fast and issue commands perfectly, ofcorse it will look perfect and superhuman. In addition, in a sense, when AlphaStar selects a group of units it (most likely) selects them one-by-one with a shift or ctrl (never played Starcraft, don't know the bindings), so you could argue that the humans have more tools

@@danielb270 During the matches against TLO and MaNa, AlphaStar interacted with the StarCraft game engine directly via its raw interface, meaning that it could observe the attributes of its own and its opponent’s visible units on the map directly, without having to move the camera - effectively playing with a zoomed out view of the game. In contrast, human players must explicitly manage an "economy of attention" to decide where to focus the camera. However, analysis of AlphaStar’s games suggests that it manages an implicit focus of attention. On average, agents “switched context” about 30 times per minute, similar to MaNa or TLO. Additionally, and subsequent to the matches, we developed a second version of AlphaStar. Like human players, this version of AlphaStar chooses when and where to move the camera, its perception is restricted to on-screen information, and action locations are restricted to its viewable region.

I bet they do!

There's no automated scenarios or so, it's a self learning system, so it's just doing what it picked up itself.

@@Quickshot0 i doubt about that

@@stkfr They release science papers on how they've made the various alpha series AIs like Alphastar, it's always been via self learning or guided self learning. But in either case it's ultimately learning how to play itself. It's in that sense very different from old style AIs that have pre-programmed reactions.

Goblin did it at NW

@@KugasKuoska that was 1 map though. and it was a pretty weird game that began with a cannon rush.

he is for his children. :)

While no one will dispute that an AI with unlimited APM will play better, no one is impressed by that either. The whole point of this project is to improve general AI development, not to win in Starcraft. It's more important to win with less APM.

Certainly. If we unrestricted its APM it would win even if it picked the stupidest strategys possible. We can always make it win, the aim is to make it win with the lowest APM possible.

Re "We can always make it win": You are really underrating the micromanaging of Alphastar here. There was no AI who could successfully take on pro SC players before, even with unlimited APM.

@@Coskunn I was referring to unrestricted Alphastar. I'm saying we have already made it good enough that it can now defeat any human who ever lived. Now we are beginning to add restrictions to it such as APM to make it even better. Basically tying one of its arms and legs behind its back so it has to learn how to win with what it has left.

The goal of AlphaStar is not to win at StarCraft 2. It is to learn what it takes to win at StarCraft 2. To that end, any behavior a human cannot perform is detrimental to the goal and a wrong answer. We are looking for Move 37s, because it's from such moments that we suddenly have breakthrough insights. Like oversaturation.

Depends. In a lot of ways alpha star is very impressive. In other ways it highlights just how far AI still has to go. The attempt at a wall off is a good example. You'll never see it in a pro game because it makes no sense. It's like it built a wall for wall's sake without understanding what it's for.

@@heroclix0rz I wonder if it does not wall off so the imo & disruptors can go in and out .... maybe it never realized that you can drop a pylon to wall off during the early game, then destroy and replace it. Just guessing, because it seems strange, I make better walls than alpha star, and I'm still really far from beating serral on a lucky day.

​@@heroclix0rz can we be sure it makes no sense at this stage? If it consistently does it 'wrong' as is claimed, that strongly hints it's no accident. The partial wall obviously helps it in some way, otherwise it wouldn't have learnt to do it. Would be interesting in a A* vs A* matchup, to force one to create a proper wall artificially (the game played to completion for both possible outcomes), then measure the difference in outcome over multiple trials.

@@Lee-pf6od The probe micro is ridiculous, so it was no reason to correct bad walling behavior. It clearly can make walls and chokepoints properly where it expects there to be more army fighting

@@Lee-pf6od You could also specifically teach it proper wall placement, inject those agents into the process, and see if they're validated or not. Basically, if the AI can consistently beat human players, we can feed it specific strategies and test them against itself in optimal conditions.

Yes, that would've been more value than the disruptor hits it actually got

Same

Yes, I think so :) I think this version of AlphaStar is either the same or similar to the 3 bots (Terran bot, Protoss bot, Zerg bot) that are all AlphaStar bots who reached Grandmaster on the European ladder. The Protoss bot got an MMR of something like 6,200 o.O

@@brofenix it's different agent with different playstyles even in the same races.

That would be really surprising if it can go that far, but on the other hand, what's more surprising would be such a crucial mistake on a basic strategy which is walling. Well spotted then

the use of deception by an AI has so many implications it's almost too scary to think about.

Maybe it left the wider opening in the wall so that it can easily move wider units through it (Archons, Immortals, etc.)?

@@jonnysith It's not that scary because the AI learns optimum strategy without "understanding" why it works. It could've simply been that over tens of thousands of games, human players showed that the APM they sink into the harassment is statistically a net positive for the AI. So the AI doesn't understand that it's laying a trap, it just understands that it gets better results with this type of wall structure.

@@vladenache6937 Alphastar did not learn from humans ukposts.info/have/v-deo/oqSdpo2Zo4WqsWg.html

Yes it's apm is limited to "22 actions per 5 secs" dunno the math on that though. They also state "with peak statistics the apm is substantially lower than those of humans." And "A* interacts with the game through a similar camera-like interface, which naturally imposes an economy of attention, so that A* chooses which area it fully sees and interacts with. A* can move the camera as an action."

They did in their AI iteration tournament. There were three exploit builds and three anti-exploit builds. Kind of an anti-anti- build.

Who is a new viewer tho

@@JacobSimonsen people who are interested in deepmind but don't know much about starcraft. it doesn't really take a genius to make that connection you know.

@@darkphoenix00001 so you take a 11 min game and make a 33 min video

@@JacobSimonsen p much yeah

Id argue that the deepmind AI ppl go watch other educational videos or play the free game yourself. In terms of president this video is extremely important to the community and it is a real shame it has been used as an educational video. This is like the superbowl or olympics so no it doesnt seem appropriate to pause the superbowl and explain what the lines on the grass mean or what shape the ball is.

It might not be possible. This algorithm probably takes up way too many computing resources. It wasn't designed with computational efficiency in mind.

@@davidliu4134 Training takes big resources but the trained network doesn't

@@seol1500 that actually makes sense. hopefully we get it one day.

@@davidliu4134 Makes me wonder if the GPUs installed on gaming rigs are completely saturated and for those that want an extra challenge could utilize local GPUs for enhanced AI matches. Of course, only some GPUs could support this and it'd have to be tested. Maybe not ultra graphics settings possibly, IDK.

Obviously it's all speculation, as these things are a kind of black box. But I think that Artosis was referring to the total amount of workers including gas miners so that would mean still 16 is optimal. I think AlphaStar is overproducing probes so that when it expands it can transfer the workers and get its economy up faster on the expansion. I was surprised that they didn't mention this at all in the January cast. This was actually pretty common practice in the early days of SC2 and there was a lot of discussion about whether it was more or less efficient/worth doing, for some reason I think it fell out of favour, but I think it might actually be more effective. It also gives you a bit of a buffer in case you lose some workers, you don't lose any mining time while replacing them and maybe it overproduces on purpose because it expects to lose at least a few workers in any game in general. I think we can actually learn from the AI here. In short, I don't think it is oversaturating because it is more efficient to mine with that many workers. It is oversaturating for other reasons. (worker transfer and worker loss replacement)

Early on, maybe, and it's a huge *MAYBE*, but later on it was just a straight up bad play.

​@@veidro Agreed. Will be interesting to see what its walls look like in a future version, though ordinary full wall-ins are the most likely option.

That's very possible. The interesting thing about machine learning is that ONLY KNOWS WHAT WORKS, it doesn't know why. So maybe it did that a few times and got an advantage and won, and just kept doing it the same way.

​@@ClayJitsu And to add to that, it might associate putting many buildings in the front with a full wall-in which is usually not at all what you want. Or making the gap too small for some units, etc.

ClayJitsu you also have to take into consideration that it took alot of things from human players, which, sometimes make absolutely no sence at all. For exmaple in some earlier games, we could see Alphastar attacking the rocks below the main ramp, which makes absolutely no sense for an AI to do. So this could have just been one of those things implemented in its plays.

AlphaStar never surrenders a match (as it says in the game chat right at the start). Any GG is therefore initiated by the human.

I assume that isn't in the works. I would guess that it is possible... but my god the amount of time that would take to train.

No matter how often and how well DeepMind explains that the “omniscient” Alphastar actually has almost no discernible advantage, SC2 fanboys disagree with not only the genius team behind Alphastar, but also the results of the two differing POVs. If they wanted an advantage, then why not keep the EPM/APM at 0ms delay instead of 350 ms delay (slower than pros)? During the matches against TLO and MaNa, AlphaStar interacted with the StarCraft game engine directly via its raw interface, meaning that it could observe the attributes of its own and its opponent’s visible units on the map directly, without having to move the camera - effectively playing with a zoomed out view of the game. In contrast, human players must explicitly manage an "economy of attention" to decide where to focus the camera. However, analysis of AlphaStar’s games suggests that it manages an implicit focus of attention. On average, agents “switched context” about 30 times per minute, similar to MaNa or TLO. Additionally, and subsequent to the matches, we developed a second version of AlphaStar. Like human players, this version of AlphaStar chooses when and where to move the camera, its perception is restricted to on-screen information, and action locations are restricted to its viewable region. Performance of AlphaStar using the raw interface and the camera interface, showing the newly trained camera agent rapidly catching up with and almost equalling the performance of the agent using the raw interface.