Hello Beasty! I'm not a Deepmind employee but I'm pretty much into A.I.. Just so that you know a bit, they don't actually program everything it does. Believe it or not, the A.I. learns all you see by itself. There is some "coaching" sometimes but it mostly learns as humans, by playing. Of course their analysis is so much deeper that they are able to get good so fast. Anyway my point is, if the A.I. is not walling off properly or doing some weird shit, this just means it has never encountered a situation in which this posed a problem to it, so there's nothing to learn from that. The A.I. doesn't recognize a bad performance for things that don't play a part. For the A.I. there's not "what if...", it's a compilation of all the things that had happened to it and that's it. This also explains the weird scouts, it has seen timings and is just scouting that, it's not scouting for all the builds we know but for the things it has seen that caused trouble. The A.I. might also do "random" things just to test if it can perform better by taking different approaches.

@zagoskintoto

5 жыл бұрын

I'm sorry, I just said "deeper" to put it in simple terms, but I wanted to refer to the fact that even if we, humans, have the same or even more memory capabilities than machines, they can "recall" things perfectly at any moment, which is something we just can't decide all the time to do. Of course my description of an A.I. is inaccurate but it intends to provide insight to those unfamiliar. Most people think A.I. is some kind of genius logic that has every pattern in some kind of circuit made out of plasma that expands magically and their brain is blue.

@HiroTeaShi

5 жыл бұрын

Thanks for this comment. It was driving me insane hearing Beasty complain about Deepmind needing to program something into Alphastar.

@melodiepinkie1583

5 жыл бұрын

+1 And I just want to add, as a result, the protoss AI is more Advanced since it has more training… But in the next months, the others AI could be at the same level!

@74URS74

5 жыл бұрын

didnt they say is similar to q learning and it just learns more and more it plays itself. Im gonna assume protoss ai is the best because they began with that race and has probably been simulating for the longest time therefore it is the best, the rest might just need more time.

@gecgoodpasi1654

5 жыл бұрын

Probably the terran ai didnt play as much as protoss either so thats why its scouting and walling feels clunky

Dude, that depo wall has a gap so that the marines can get in once it raises them, and no zergling can get out :D 200IQ

@andreab6842

5 жыл бұрын

no this means that the AI copied from real player, but didnt understand actually the reason you raise the depo

@shinmalphur9214

5 жыл бұрын

ai logic: you think i'm trapped here with you? NO! you are trapped here with ME!

@executorarktanis2323

5 жыл бұрын

@@shinmalphur9214 I don't know but when I read your comment it was like best was saying this line

@BrandonZeeb

5 жыл бұрын

Moves that bring the AI closer to a win state will get priority, the crappy wall offs and structure placement could be just “noise”, or attributes of the game state that have less utility.

@andreab6842

5 жыл бұрын

@@shinmalphur9214 rorschach

Regarding the "how can it click x/y/z" question: AI doesn't 'click' at all; it just sends the relevant command event and positional data to the server. The fact that it can command a CC to an invalid position isn't actually all that surprising; it just means the on-command check for positioning is part of the client-side UI (which of course the AI doesn't have). The developers probably decided it wasn't worth verifying server-side on takeoff if the server has to do it on landing anyhow, and just left it in the UI instead.

@Galahad54

5 жыл бұрын

the robot API is different, and doesn't need groups in the traditional manner (or clicking). So it's a bit faster going through all that, which leads to EPM very close to APM by human standards. Also a sure way to ID a bot (or hacker).

Guys, the timing of building placements and of potential problems occurring because of this are widely separated. Similarly for scouting (or lack thereof) to some extend. Such correlations between events separated by longer intervals are notoriously more difficult for the AlphaStar AI design to pick up. Beastly is not an AI expert, but he in my opinion, he does a stellar job (and yes, pun intended) at pointing out what clearly are mistakes by alphastar and calling them out as such. I would very much appreciate his coverage if I were an alphastar developer. One last thing: if you've been following AlphaZero as well (whose design is different, but share many behavioural aspects still), you will know that the AI shows most precise play in cutting edge situation where the game outcome sharply depends on which action is picked within the current search space. Conversely, when the AI is put in a situation of almost certain win or loss, the evaluation of available actions become degenerate and this induces seemingly random play. DeepMind has repeatedly shown little concern for that "problem", and understandingly so as they are working on building a winning agent, not one that should pass some kind of SC2 Turing test by showing human-like play at all times. So I would encourage Beastly to mostly consider assessing AlphaStar actions during crucial win-or-lose moments.

The question still lingers Can deepmind actually deal with liberators The AI nightmare

@Miranox2

5 жыл бұрын

All of Starcraft 2 is an AI nightmare.

@Galahad54

5 жыл бұрын

DeepMind AlphaStar has played hundreds of millions of games against bots, and maybe 200 against humans. It's like only playing on a LAN in your part of town, and suddenly meeting other players. Of course, this also implies that your hometown LAN is in Seoul, South Korea, so there's that.

I find Beasty's analysis a little bit unsatisfying because he keeps going on about building placement and overlooks the fact that this AI is dominating humans. Maybe tetris is not as important as you think it is?

@yamchadragonball6983

4 жыл бұрын

I think the AI has learned that macro is the most important, followed by micro and its just not gotten around to figure out it can achieve an advantage with building placement yet. I would assume that it starts by learning whatever swings the game most and then slowly smoothing out less important issues such as Sim City'ing correctly.

Terran AI knows better - "lets place building like bronze player, sure it will be suboptimal but the benefits outweight cons, enemy players seeing that will assume im rather shit and easy to break which will give me tactical edge"

Maybe human players waste time placing buildings in aesthetically pleasing positions.

@ieaturanium574

5 жыл бұрын

it is not just aesthetics, the ai just lacks any logic behind building placement and wins with brute epm macro/micro

@RBAWintrow

5 жыл бұрын

@01:49 it selects the Barracks even though 99% of the barracks model is offscreen. Without a hotkey.

@Miranox2

5 жыл бұрын

@@RBAWintrow It can even select things that are not visible sometimes. It's likely the bot is playing with 4:3 ratio screen, so the replay screen wouldn't show everything the bot can see.

@davidbodor1762

5 жыл бұрын

@@RBAWintrow It might have internal hotkeys? Or maybe its camera is slightly different than ours...

@yamchadragonball6983

4 жыл бұрын

@@davidbodor1762 It doesn't see the game in the same way that we do. The devs tried to explain this during the show matches against mana and tlo.

AI: I didn't just let the lings come through freely, no. INSTEAD, YOU ARE TRAPPED IN HERE WITH ME. XD 16:00

You don't understand. That depot raise in the first game was a reenactment of the scene with Rorschach in the prison in Watchmen. "I'm not locked in here with you, you're locked in here with me!"

There isn't just one "AlphaStar AI". There are many *different* AI agents, each with their own personalities, if you will. They learn independently of one another. I am fairly sure that each 50-game account is its own agent, being trialed on the ladder to evaluate its performance.

those scattered depots you usually see at the 3rd and 4th, but they were pretty effective in not allowing zerglings to get a surround

To me, it looks like the Terran AI is much younger than the Protoss AI. A lot of these mistakes are "This choice has not yet changed the outcome of the game" thinking. I don't think AlphaStar believes that its mistakes are correct moves, only that these particular mistakes are irrelevant decisions. As it plays more games, it will begin to find these mistakes. Did they use a more mature/experienced Protoss AI as the base for the current one?

@peterhardie4151

5 жыл бұрын

As a 100% salty Terran, I Think its that Protoss is easy.

@jamalyeboah4322

5 жыл бұрын

Deepmind started with the development of the Protoss AI far before the Zerg and Terran ones and it was what they originally showcased. The Protoss AI had more experience and training time that the other alphastar agents so most likely it’s had more time to work out its kinks.

@StargazerWindwings

5 жыл бұрын

@@markkunitomi7901 > the Protoss currently playing is probably completely independent from the one previously used OpenAI (another AI company) has developed technique to transfer the experience from older neural networks to the newest ones. This way they can continue training the network in parallel with improving/fixing and expanding it. Maybe Deepmind can do the same.

I think I understand why it's bad with T and not with the other two races. Alphastar has an incredibly hard time predicting the future, in order to be a good Terran player you need to predict where the Enemies army will be and siege tanks ahead of time build walls ahead of time and generally predict the future. Also I think the fact that Tara matchups are so drastically asymmetrical in strategy affect it's play style and it's trying to go for good all-around builds which she can be easily abused.

Did you notice how the ai pulled the mule from the mineral line at 37:34?

@DavenH

5 жыл бұрын

omg that's hilarious. The action has zero cost at that point cause it's gonna die in half a second, but it doesn't know how to prevent its death.

@NoOne-tg3xb

5 жыл бұрын

@@DavenH no it usually does that to not waste minerals, some pro players do it because the mule doesnt have time to return the minerals it mined, therefore it would lose some minerals from the mineral patch.

@DavenH

5 жыл бұрын

@@NoOne-tg3xb wow that's a pretty tight play then

The depot placement is 300iq You just dont understand it. In the end if there was a wall the zerg would have go back home. But what happened here? He trapped the ling so they could not run back home neither attack yhe marines. In the end the zerg lost all his ling for 3 useless scv. Also the "weird" tank placement was here to wall off the entrance between depot.. just open your mind

@Galahad54

5 жыл бұрын

Exactly. If you're starting your all-in, losing a few SCVs while trapping the zerglings away from defense worked here. Also, the AI does not do a single evaluation (win-lose). If it tried that (Google did try that) it would take hundreds of millions of games (they tried that for hundreds of millions of games) to beat the Easy AI (actually, after hundreds of millions of games, it still couldn't beat the easy AI). I just re-read the Google paper on that. Also, I'm writing my own AI for Starcraft 2. Look for Krillin (and more to come) on ai-arena.net

@simeonpolet1307

5 жыл бұрын

You could do the exact same thing with a full off wall, and protect your economy while attacking. I don't see any IQ play there. The IA just never face a situation were it was a problem and that's why it's doing this.

@tgaudlol

5 жыл бұрын

@@simeonpolet1307 it walled with a tank. Its a way to wall off. And maybe for the ai having an entrance large enough for a marine but not for a roach was a good thing.

The first game supply depot were used to wall zerglings in his mainbase. Its like a mice trap, you bait zerglings in with cheese aka unprotected SCV's and then snap the wall goes up and they cant get out.

do you remember the DOTA AI where the AI was making plays in order to bait humans??? I think this might explain the bad wall off. The AI has noticed that if it has a bad wall off, the human will over commit to ling flood maybe, and that leads to higher percentage of wins if it holds??

@Atlantis357

5 жыл бұрын

It didnt train vs humans, only vs itself

@thomasdieter9302

5 жыл бұрын

@@Atlantis357 Yeah but with knowledge based on human replays

As others have pointed out, I also feel like the criticism against the AI is a bit too harsh and unreflected. You can call it bad compared to human professionals but the whole project is academic research and in this field, you have to compare the performance to previously proposed Starcraft AIs but in fact, there are any. So this work, although it's the very first of this kind already shows incredible good performance which will most probably drastically improve with further training and optimizations in the neural network model. Same as AlphaGo improved from about-human level to absolutely super-human, 100 to 0 beating level, the Starcraft AI will also reach this point, probably within one to three years.

@ChazAllenUK

5 жыл бұрын

It's not bad; it's focused. Clearly precise building placement is not the most important factor in winning games.

"its hard to program", OK, I understand not everyone has masters in C.S., but it's not hard to understand that these kinds of AI are not programmed. They learn to play the game from scratch. No human "programs" in the way AI plays. Its completely self taught how to play the game. People only program the learning algorithm that is NOT custom tailored for SC2, it can learn any game

I suspect that when alphastar was positioning the tanks very defensively vs zerg it was because the win condition of that game was mass tank pre-hive and hence it valued their safety enough that it didn't care if they weren't defending optimally. Additionally I find it interesting that alphastar chooses not to scout with terran (the easiest race to scout with) as it clearly doesn't care what the opponent is doing.

@DavenH

5 жыл бұрын

It could be there are indicators too subtle to pick up by humans, that give it clarity on the opponent's build. It's hard to know

Just wanted to say how great this semi series is; you are trying to give a logical explaination to every ai's move and it alawys makes a lot of sense which I find amaizng

This Bug with cc occurs in different situations in normal games for example if u are placing cannons with rapid fire they sometimes get placed in the same spot and the error shows up when the probe is Tryin to build cannon

@ETpwnH0me

5 жыл бұрын

That only occurs because of lag making the server place the second cannon down before it has set the first ones build location. It's totally different than building too close to the minerals because that's already set as an impossible spot.

@ekothesilent9456

5 жыл бұрын

John Smith you can actually test this very high ping, the mineral locations are not preset even though they seem to be they still need to be updated to the server as eventually they can be mined out and their spaces can be built upon

the first game was rly strange in many ways but those two sensor towers, they got me good lmao

@lloydsloan1349

5 жыл бұрын

ikr?!

I think it popped the factory down when it saw the second hatch in the main, because there weren't mutas and so it was safe bet to go into more tanks to deal with the influx of units

Absolutely no one - *ling run-by* Alphastar - “I’m not locked in here with you... you’re locked in here with me.*

As a researcher of AI, I'd like to provide some inputs to my speculations of why Terran AlphaStar messes up the building placement and depot shenanigans, and my reasoning behind why AlphaStar is behaving like what we've observed. [DISCLAIMER: My main interest of research is in computational linguistics, and I'm not very experienced in reinforcement learning, which is what makes AlphaStar happen] So the big issue with games like Go and especially SC2. So AI that plays games traditionally uses a technique called Q learning, which builds a table of (game) state vs action called Q table. The number inside is called the Q value and signifies how good of the action is of that state. With each iteration, you update the Q value in the Q table so that the AI becomes better. With a game like SC2 and go, there are way too many game states to make a Q table. There's a deep learning variant of Q learning which can handle huge game states, but it doesn't work well enough when the order of magnitude of game states is as large as go and SC2. What AlphaStar is doing is much more complicated, but it's still generally following the rule of the more game state there is, the harder it gets. When we're looking at Protoss and Zerg, buildings don't fly or have add-ons. In other words, the mechanics surrounding buildings is a lot simpler. This means that the AI is more likely to explore most of the states (or situations) that is likely to happen in a game. For Terran, it is much more difficult, as having depots being able to raise and lower and affect pathing, and buildings may optionally add on a 2x2 add-on which could not be moved is complicating it a lot. If you go to lower leagues, you will also see a lot of people messing up Terran building placements, but not Zerg or Protoss building placements. This can also explain why the Protoss AI is in much better shape than its Zerg and Terran counterparts. In terms of unit compositions and strategy, the Protoss race is really easy. In the early and mid-game, zealots plus immortals (maybe mixing a few archons in there) are very strong and counter almost all ground-based armies, and stalkers with decent micro are also really strong against most air harassments. The other technical units are specifically for specific scenarios, and you will make a few of them to add to the army depending on the opponent's strategies. For example, when the AI sees a lot of high DPS low health hydras (glass canons), it can add in a few colossi or disruptors for quick AOE clearance, but when it sees mutas, it can simply add on more stalkers and research blink. For Terran, there are two main paths: mech vs bio. This creates a lot of variety and is difficult for the AI to explore. Some of the most-used Terran units also really depends on the sieging and zoning mechanic (siege tank, liberator, cyclone). It also needs a lot of strategical thought on unit placement (sieging location choices, bio pre-split etc). Many units also achieve the same purpose. For example, to counter air armies, you can either make vikings, or more marines, or a squad of cyclones. These all add to the complexity of army movement and unit composition choices. The AI would need to be able to explore a lot more of the possible game states to be able to perform well. The army is much less A-movable and thus more complicated to use. Since the AI's "thinking" is in terms of probability of what move to make next, it's going to struggle to choose between bio/mech and the different tech choices. For Zerg, the opening is pretty straight-forward. you always start with lings and go either bane or roach, and then you go to tech like muta or hydra, and then you go to the late game which opens up a lot of choices and you can choose depending on what your opponent is building. This means the AI doesn't need to scout much because it's going to have a strong army by simply sticking to the roach ling A-move army. It will be able to beat its Terran and Protoss brothers by crazy macro and overwhelming the other AIs (which don't do weird things like building a bunch of void rays). This is most likely because of the insane larva mechanic which requires the non-Zerg opponent to harass and kill drones to keep up with the macro. The AI most likely didn't harass that well when it plays against itself, so the Zerg AI would believe that roach ling all-ins are the strongest strategy and would tend to scout less and all-in much more often. I'd love to know if anybody sees anything that counters my speculation. I'd come up with some better explanations for that.

@Newbtuber

5 жыл бұрын

I believe it is pretty spot on which how I am viewing the AI. I am curious at how the AI will do if you hide tech such as a dark shrine or simply just building your entire base outside your base (as protoss). I find it hard for an AI to understand that the human player is hiding tech and how to formulate a strategy against it. Will it think that the enemy base is weak so it will try to end it or start making cannon or defenses preparing for DT. It seems like a paradox for the AI since he would As Beast said in the video, he said that the AI did not scan or scout for Hive timing. Which mean it either doesn't care or it is attacking the zerg at a timing that would kill him if he had the hive in which case it doesnt matter. I really just curious how the AI function with limited information which is why I never believed an AI could play starcraft. You also forgot to factor map into the equation. Wouldn't different map mean the AI has to relearn how to play on that map

macro is key! You can play stupid but if the macro is on point you win. Something what we humans can learn about this

I think the AI saw the hydras by calculating the time of hatching of the eggs. Different for every units

Tbh, I think that not all the AI is on the same level/training. That is, there may be multiple accounts with different AI training levels to see how it places on the ladder.

@ganonk79

5 жыл бұрын

Their initial demonstration of AlphaStar was actually many different bots, each playing a different match. They could very well be doing the same thing now.

He didn't scout 3rd cause he watched the second queen hatch which is usually a timing where you get exactly 300 min when ur pool finishes and you can either spend on base or both queens, seeing the queens early maybe gave it the info it already needed

it may not build double engineering bay on time, but has double sensor tower :D maybe it was just testing, if it spots double the amount of enemies

3:50 the cool stuff about this (AIs in general) is that you dont program it to play, you program it to learn, so building placement is something that probably wasnt enough of a problem or was overshadowed by superior micro in lower leagues, but will eventually get better

When you paused the replay I had a heart attack LOL

There should be an alphastar AI that anyone can play against in game with different difficulty levels restricting the APM more or less. I feel like the more this thing could play against humans of all levels the better it can get. Plus I just wanna play an AI that can think. I remember playing SC1 as a kid and keeping the AI alive after I had won the game, hoping it would try to rebuild and keep trying to win. It's nice to know that there is an AI with heart out there.

The AI didn't "click" anywhere. It sent a packet with the coordinates for the landing. It just means that the LZ target check is done on the client (which makes sense) instead of the server, and the landing is allowed or denied by the server (which also makes sense).

An important point is that the AI learnt to play by playing a lot of games. However, it is actually different versions of the AI playing against each other. Deepmind can choose any of those "agents", trained after any "number of games". It does not need to send the GM level AI.

About the CC landing, SC2 checks for whether you can execute an action both when you issue the command and when the action takes place. In maps where you set vision to unexplored (like in the campaign), you can command a base to be built near a vespene geyser if you haven't seen the gas but has enough vision up until right up to it where you want to place your base, then it won't build not when you gain vision of the geyser but when your worker reaches the destination. In high latency games, you issuing a command to to a unit that gets disabled after you issued the command but before your command reaches the server will also not actually execute the command despite you issuing said commnad.

What happens if you pause the game vs Alphastar? Like pause and walk away from computer for 10 hours type of deal?

Those supply depots kept the lings _in_ ! It was a clutch move indeed!

That was crazy that you found that "bug." Fascinating.

what happens if you use pause against it?

I mean, that Wall and runby defense is very conparable with the last PvP Alphastar. Remember how that Alphastar also never (or very rarely) bothered to Wall to defend against Adept harass? It seems Alphastar is willing to trade SCVs for lings, same as it did probes for adepts.

Did it "click" at all? I mean in the normal way. Or it is somehow different like selecting a pixel on its screen where the "action" will occur, but no "click" happens, only the command from the game engine.

13:10 So that's why it left holes in the wall

The CC landing next vespene geyser looks like a bug. It happened to me but in coop mutation using Commander Nova. I ordered my SCV to build CC not knowing it was beside a geyser. When my SCV reached the destination, obviously he said "There's something in the way"

Supply Depot wall placement was brilliant. In a linear wall off, if there's one massive breach you're fcked, ie the wall completely protects you or not at all. If you stagger them and leave very small gaps and layer them you create a multi layered maze. The objective here is not to PREVENT a breach but to SLOW it down.

@dannygjk

5 жыл бұрын

Interesting.

no wonder they didnt show alphastar as terran. be hilarious seeing stuck tanks & bad building placement, no frontal wall etc... xD 2 sensor towers cuz

at 10:50 in the first game, AI builds 2 supplu depots while at 200 max supply already. is that a GM thing, or a weird AI quirk?

43:26 *commits sudoku* XD oh man it took me a second to realize you misspelled it... then it took me another second to realize... wait a minute.. oh man i haven't laughed that hard in a while X'-D ty kachinsky

So, the last time, when Alpha Star played as Toss, it was explained that the computer learned by playing against humans. This implies two things: 1) Every action, placement, and timing, in short, every choice the computer makes, is not the computer's original thinking but the selection of a choice a human once made, and that the computer thinks best fits the scenario. 2) Each choice has a value, a score, attached to it. The computer picks what choice has the highest value, and does it. But these are not weighted like we do as humans. The computer is using a law of averages. And that is not to say that humans learn any differently. On the contrary, that part of it is the same. The difference is humans prioritize and emphasize data points that they deem more valuable, while the computer uses ALL of the data. This greatly reduces the work load for the human mind, and increases the rate at which we optimize, but it also means we ignore and discount possibly viable solutions. Once the computer plays enough games, it will have tested out all possible solutions to any possible problem, and thus, always make perfect decisions, and then no one stands a chance. But it is still gathering data. Humanity stands a chance for now. p.s. This also points to the difference between humans in any competition and computers: feelings. Morale plays a significant role in all sports. It can also be preyed upon; baits, traps, tricks, switches: a person can play another person. The omniscient A.I. doesn't have feelings, and thus cannot be played.

@johnuferbach9166

5 жыл бұрын

afaik it first watched human replays to start out and then wemt on to play against itself developing own strategies... there is no predefined decisions it's choosing from, neural networks have some amount of inputs, one or several layers of nodes with weigths and outputs, which in this case are the commands it issues to the game.. and what it does for each input is decided by the nodes and their weigths... and training means having parameters that define how well the ai did, so the training algorithm can pick / merge / evolve the ones that did the best under those parameters

when you are placing buildings fast and using shift, sometimes is shows that the buildings are overlaping. maybe its the same issue with the commandcenter. because it also the case that the buildings will not execute the command

Every time you pause it and get inquisitive I can’t help but laugh 😂

As others have said, DeepMind doesnt "fix" their AI, it has to learn on its own, to elaborate why i think it hasnt learned good walling yet: I think the problem with building placement is that it is such a complex thought process behind it. Modifying your game world to positively affect the potential unit flow in/around your base is not an obvious mechanic. For the AI to learn something easily it needs immediate and clear feedback on its actions. Placing a building and preventing a rush might not have happened for the AI often enough to make the connection between the two distinct events, especially since the actual placement position requires a basic understanding of fluid mechanics (for humans common sense). So if the AI had a button "Block ramp" it would surely do so, but using the ramp position on the map to calculate building coordinates for a tight wall would be something even humans would struggle with.

@wildwest1832

5 жыл бұрын

IDK if you think about how a human learns its mostly from other humans too. Actually how a neural net learns is actually very human like. I didnt know how to wall off or the concept of it until I saw other people do it, and thought oh that might help me stay safe. An AI could learn it the same way, but maybe the question is does walling off correlate with winning or is not as important as our human intuition tells us it is? If the AI never walls off maybe we are wrong and its right. Maybe controlling unit flow is more useless in practice than you think it is.

You have to remember that there is no single AI, there are several that compete in "training" leagues. So what we are seeing is just a few of the possible AIs that the AlphaStar process has produced. And they are all bound to play very differently.

For anyone more interested in the alphastar deepmind project i believe this is an interview with the person behind it all. It's a long interview but it covers quite a bit of ground including the fact that they have a internal battlenet where starcraft agents/races are basically competing against each other. "Oriol Vinyals: DeepMind AlphaStar, StarCraft, and Language | Artificial Intelligence Podcast " kzread.info/dash/bejne/fZmY1pSoopzenNI.html

From what they said the ai doesn't actually see just a small portion of the map, it sees the entire map at all times (with fog of war included of course) which could explain why it doesn't have control groups or camera locations

@RodelIturalde

5 жыл бұрын

Not anymore, this was the case for the protoss games against Mana. Then the creators limited the view for the AI to be as large as 1 screen, a nd then force the AI to move the screen to view other parts of the map.

It did not raise the depots to keep out the zerglings accidentally keeping its own marines out - it raised the depots to trap the zerglings and kill them all. Seriously though, I think it is producing tight corridors and values them above own units ability to move even. Same goes for the building/addon/turret placement. It deliberatly makes really hard pathing for melee units. One-tile-gaps are everywhere. The AI would've had no problems recognizing it was already too late to raise the wall - it did expend its valuable actions anyways because it wanted as many tight spaces as possible. I can see that evolving as a winning strategy against zerg and protos pretty early on. By this time in the learning cycle it is probably one of the more basic things it just does anyways every game not even considering other options very much anymore - along "mining minerals raises my chances to win". Have zerglings in your base? - Tight spaces are paramount. Raise all the depots, land all the buildings. It is the more apm economic version of having to control your whole squad of marines and scvs perfectly. Probably the same reason all the depots are at the natural - most encounters with zerglings and zealots would be there instead of inside the main base. So in a sense all the guys refering to the "you are trapped in here with me" line do have a very good point.

IMO the supply depot placement is supposed to counter the late-game. Ultras can't get through the walls, but ling and blings can, which the marines can cut down on the other side of the wall.

@3:49 They don't "program" the building placements into the AI. It just learns it, and if it does end up sim citying its own troops in in longer games, its not going to happen too often and isn't too penalizing.

At 1:49 there’s an interesting barracks selection, no hot keys and barracks is so so so slightly visible

3:40 remember that the AI isn't "programmed" it's just an empty (insanely big) matrix that maps inputs to outputs in the likeness of a real human neural network. Maybe its objective function isn't correctly reinforcing getting stuck in the base as something very negative. EDIT: The cc landing thing might be issuing the command like one frame before collision det, or the mouse input is not through the UI but straight to the api and is missing a check.

Maybe the AI starts with the unexplored portions of the map as completely black, instead of greyed out like the human. So to it, there was no geyser to block the CC placement, since it had never explored there.

I think about the CC landing next to a geyser, the AI could potentially land it. I would think that AlphaStar is not interacting with sc2 the same way as we do for sending commands, therefore if the AI issue an impossible command, it might still execute. In one of the game, it tries to build the CC at the natural but it has the ramp walled off and did not lower the depot, result it built it on the high ground and was a bit confused which made it lose time and got supply blocked. At 37:40 . In the same game at around 39:40 it issues the order to the reaper to go at the very top of the map where it's clearly unpathable. And then at 40:02 it tries to go at an unpathable high ground area. I think at some point the AI should be able to know the ground pathable grid of the map so it doesn't explicitly tries to go to weird spot. Or, it needs a way to understand that areas with doodads and stuff tend to be blocked. And, how does it interacts with air blockers? I'm very intrigued to what other interesting situation could lead to unusual stuff and even more... to impossible things. Like, what if a viper abduct a viper that is abducting or whatever. I mean.. that won't work, but what have we not tried that AI could use? Like, what if it finds out how to stack tanks by facing eatch other? What if it mines with Nydus efficiently by knowing which drone within the nydus has minerals? The AI is trained by itself, but imagine if you promt it with some cue about what it could do so it trains in that direction and come up with very powerful strat ? How fast would it takes it to creep the whole map? How fast could it just max out on drones and optimse expanding early and just mass drones? And what if they give it limitless apm to be able to optimse mining and everything as much as it wants to? Maybe even be able to come up with precise math and timing about how fast you can get X unit. It's just crazy. I want to see more of this! It's so fun... but scared at the same time of what it could be used in real life situation.

hmm does besty really think they program what the AI does ingame?

The Terran AI is more focused on the macro, while the Protoss AI is more focused on the game sense. This could mean that macro-centric Terran might be a more effective path.

29:11 this could be a bug in SC2's AI API. This is basically how AlphaStar interacts with the game, there's a separate client that the AI connects with and this separate client then connects to the game, and there maybe a bug with this in the AI client that AlphaStar connects to.

Terran ai not scouting doesn't seem weird to me. Just building bio safe vs everything was standard terran play for a long time. AI is taking the most effective strategy and just making shit efficiently is the best terran strat. Seems pretty simple.

ai built a second sensor tower to wall off his expansion :D

My only complaint with these videos is the CONSTANT rehashing of the same topic: building placement. We get it, you don't understand how the AI thinks.

A reason it's different I. E. Scouting third, might be because of the matches it played before and learned from. Might also be the code varying itself a little xd

IM EXITED! lets GOO

I feel like terran Alphastar is just young maru...

A.I doesn't click, it issue commands which bypass the UI clicking. That's why there's no delay like humans moving the cursor from point a to point b, it just instantly select units(multiple of them if need) at once or move from mini map screen to screen to macro instantly. The error msg and prevention of placement is probably part of the UI which is why Alphastar was able to bypass it. However that command cannot bypass the game mechanic, which is why at the end it couldn't land the cc.

@skychaos87

5 жыл бұрын

@AFfen Mann Well yea, they did implement delays so the A.I can't issue commands in quick succession. However it still doesn't issue command through the UI and there's still advantages. For example, in 39:33 you can see its selecting units and structures that is not in its screen. Humans do that with hotkey, but i think Alphastar simply does not select units, it just issue commands. You can see that Alphastar's tank is selected(have that circle around the units) and execute the command at the same instance. Human players would have to first select the units, by mousing over it and click, then move the cursor to another point and click again to execute the command. So well they did implement delay to simulate human reaction time, but it does not have to deal with UI limitations like humans do.

@skychaos87

5 жыл бұрын

@UC6CICkOx1p4m0b295UsNo4A It most probably won't land, the game mechanic takes precedent over player commands. The difference is that human players cannot issue commands that the UI prevents(i.e the red color blocks that indicate no placement zone), the A.I simply bypass that which moves the CC regardless, the placement command is then later not accepted and canceled by the game when the CC reaches the position. Its no big deal since it doesn't break the game. But there's still some advantage because of that. For example, if you cleared the creep tumor or overlord spawning creep and your units lose vision of the receding creep when you move them away. The creep will remain present in your view(without active vision) even when you know its gone after a period of time. The UI prevents you from building anything on it and you need to move your units there to make sure the creep is gone before you can build anything on it. That's 2 command issued, #1 moving to point A and #2 building on point A. Alphastar would only need to issue one command of building on point A and the unit would just move towards the area and build, as it goes near its active vision would show that the creep is gone and it will start building.

19:18 Some skynet shit right here.

Oh but the wall was never meant to keep them out....

@BeastyqtSC2 @01:49 it selects the Barracks even though 99% of the barracks model is offscreen. Without a hotkey. If I tried that the screen would scroll.

This terran is brilliant! He made a lot of small mistakes, but still wins! It means, that all that mistakes are not important. Everything what he does is important!

@Miranox2

5 жыл бұрын

The bot is still far from its maximum potential. You will see the mistakes steadily disappear in later versions.

@erbgorre

5 жыл бұрын

no, it just means that superior macro and micro at a level that is physically impossible to reach for an organic being is ultimately more important than how good your strategic thinking is.

@Miranox2

5 жыл бұрын

@@erbgorre This bot does *not* have micro/macro beyond human level. You're thinking of the old Alphastar that beat MaNa.

@erbgorre

5 жыл бұрын

@@Miranox2 this version of alphastar still gets like 8 worker kills every time it harasses with 2 adepts, or 10+ with 2 banshees, every time. it still controls armies simultaneously accross multiple screens with no misclicks and virtually no micro errors without even needing control groups lol. no matter what kind of restrictions they try to put on this thing its micro and macro will always be leagues ahead of what any human player can do, because it can interact with the game on a fundamentally different level.

@Miranox2

5 жыл бұрын

@@erbgorre Alphastar makes plenty of micro mistakes and a few misclicks. Humans can also get that number of worker kills. Beastyqt explicitly stated that *he does not see Alphastar doing anything superhuman in any of these replays*. The EPM and lack of control groups are the only abnormal aspects of it.

Awesome video beastyqt!

wait wait, beasty if you join in the game from a replay as alphastar, does that also mean you can check if the CC will landed onto the no -no zone? edit: nvm, xD nice

The AI picks from a pool of different trained agents every game. Each game is potentially a different agent so it's not going to have the same playstyle every game. Thanks for doing these videos btw. It's interesting hearing the high level commentary of a non-human player.

How do you snipe an AI??

When I saw the walling in the natural, I thought the AI had this genius idea of sealing the zergling or something that humans cannot understand 🤣🤣🤣🤣

you know its the AI, it scouts every 30 seconds for information. I played it when it was a zerg, overlord speed was first upgrade and every 30 seconds an overlord was scouting from all sides, main, natural, 3rd, then started scouting 4, overlords everywhere moving around for scouting information. Ive never seen a ladder player do things like that

As a plat player but graduated as twitch chat GM I usually try to siege my tanks in spots that will cover as much ground as possible while splitting them. So if the enemy player tries to run towards the tanks they will have tanks spread out shelling everyone

@KaizerKilborn

5 жыл бұрын

Alternatively, if you group them, they can all hit at once causing many more enemy deaths before they understand to retreat with them from tank fire.

Well, I am pretty sure that if Deepmind gives more training time on various map pool then AlphaStar agents will finally catch the concept of building placement. (DeepMind once mentioned that AlphaStar league are trained on only one map. So, optimal placement is not a big deal if the map is fixed. in that circumstance, it will just tend to memorize the solution dedicated for the map) Also the agents that are playing in EU server seem very premature when they are compared to the agents that beat TLO and Mana.

I think its pretty good AI man. Based on how it learns from games as a neural network its pretty far up :)

14:25 a two step plan on how to throw as zerg: 1. F2 2. attack move.

I guess that AI's weird moves like sending an SCV to scout and return it from the middle of the map is some kind of learning artifacts. Like it half sure it needs to scout so it does a halfway scout. My point is you can't expect AI to be 100% rational.

The machine learning aspect of alphastar doesn't care about placing buildings nicely or any of the other technical aspects that were touched upon in this video. It only cares about winning. It has a huge volume of real games to learn from and then it hosts an almost tournament like structure of simulated games to find the "best" agent which can win against the most different types of strategies. To say this particular agent is platinum league is absurd when it clearly has a very high win rate even in high masters.

Here is my hypothesis about the terran struggle thing, It plays like a game of chess, as it goes it uses building to make a "wall" in strategic defense place. Plus, this gives to the AI to scout as it goes! with the base indeed. The fact that terran building can move makes input MUCHHHHHHHHHHHHHH heavier for the AI and it learns slower. Havnig the AI doing this work slow it ability to decide about strategic unit "only" plays. The reason why it plays like that ( my humble opinion ): It wants to make the infra that build unit as close as possible of the main army so that if a unit dies it can makes a new one right off and reach its army faster as well. It's also more defensive... think of it as "wallling" the base so you have to make a breach so you can pass, and again, and again. But as you may see it does it way too stupidily ( making those wall pack of building ). That's because those decisions are made too early in the game. And it does answer you relatively about what you write to it .... pretty sure about it! Not so stupid but not smart enough...!

This is a fucking computer. Holy shit. Making hella intricate decisions. Like when it used those cloaked banshees, he didn't just attack the first thing it saw. It gained that information, then patroled to pick off a few units then retreated. It was so calculated. Imagine the campaign mode missions they can make using this lol.

I want to see an Alphastar bunker rush!

To me the AI looks like it was given hellion banshee openers as a base to learn from and while it's trying to do what Human Terrans do, it it doesn't know why it's doing it.

No scouting explained: It's been playing against itself for effectively 200 years. If you've played the same or similar opponent for 200 years, would you need to scout? No - you always pretty much know what they're going to do. This is it's first real experience against anything other than itself, and it's such a tiny little blip - I mean, it's played 100,000,000,000 million games against versions of itself, and 50 games against humans. Explains quite a bit about it's behavior.

22:50 it checks the creep with the reaper because most likely the previous one was a different agent. They all act a bit differently, remember. Even in abest of 7, you'd find it difficult to play against the AI because they would shuffle it for you just like they did for TLO or whoever was facing the protoss one.

I think it did see the hydras at 9:45, look at the egg, it can see a hydra being built there...

alphastar just need to become a streamer.

Just ran into it, first time I heard they released AlphaStar to the wild, after the matches with TLO and Mana. I wonder if it learns on the go like Leela Chess Zero does, if so, I won't be surprised to see it play on a high GM level in a few months.

@dannygjk

5 жыл бұрын

Nope they do not learn while playing in competition training is a separate process.

This AI is definitely more reasonable than the ones that went up against TLO and Mana as far as micro goes, but it still doesn't really feel like it's winning through superior strategic thinking. It's better, but it kinda just seems like it hits all its production cycles, never misses a beat macro-wise, and wins. I suppose there's a lesson in that for human players - that sometimes all it takes to win is getting good at the absolute fundamentals - but it still doesn't achieve deepmind's real goal. The thing is, I have no idea how they could correct the issue to force the AI to think strategically. The micro problem seemed like an easy fix: limit its APM... Macro just isn't like that. A computer won't ever forget to macro while it micros while any human (save our lord and savior Serral) will. That alone gives it a huge advantage that can't be easily neutered for the sake of actual progress strategy wise. It isn't all that impressive when a player simply outmacros their opponent, a-moves and wins. What's really impressive is when someone can out-think their opponent and win when the odds seem against them.

@dannygjk

5 жыл бұрын

With NN systems the strategic skill emerges from training the net(s). AS's behavior/'understanding' is not coded.

@dannygjk

4 жыл бұрын

I'm back! lol. It's not impressive to me when a human beats another human because of superior micro.