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
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 was really curious whether they would attempt moving the camera like a human. Sounds like it's still a work in progress, but very exciting! Even this isn't enough to make it fully like a human player, as I believe it is still getting numerical values for unit properties rather than having to infer them from the pixels on the screen. But it seems possible to fix that, likely at the cost of drastically increasing the training time.
The benefit of using pixels, of course, would be that the agent would become fully general. It would probably immediately work on Command & Conquer, for instance, while the current version would require deep integration with the game engine first. But I think the training time would be impractically long.
The live game that was just played was against this version of AlphaStar. Mana did win, but it was by exploiting some poor defense against drops and hard countering the stalkers he knew AlphaStar favours. The AI still looked very good and the developers claimed that this version of AlphaStar wasn't significantly weaker than the versions which didn't have to use the camera.
Dealing with perfect blinking is basically impossible, since you can blink back your units right before they die. Stalkers are balanced around the fact that HUMANS have limits to how well they can micro.
While the "skill cap" on blink stalkers is extremely high, there are many hard counters that can stop even perfect blink micro. MaNa won because he went for one these. Immortals are the perfect hard counter to stalkers because
- cost-for-cost, they are more efficient in a faceoff (resources)
- immortals are space-efficient dps (damage per second) in a battle. In a given battle, an army of 4 immortals is far more likely to all be in range of an enemy and doing damage than an army of 8 stalkers bumping against each other trying to get to the priority target
- immortal shots do not have projectiles, but are instant. No matter how perfect your stalker control, once an immortal targets a stalker, it is guaranteed to take 30+% of its hitpoints in damage.
The last point is very important. Once MaNa had 3+ immortals, even with perfect blink micro, a little bit of target fire and timing micro on MaNa's part allowed him to slaughter the stalker army one stalker per volley, while it takes them longer to clean up the immortals (especially with shield battery support).
Another thing glossed over in this discussion -- AlphaStar did more than classic blink micro. It did a very technical maneuver (the casters briefly allude to it) of triggering the barrier on one immortal with a single laser, then focusing all fire on an immortal whose barrier was already down from a previous iteration of this tactic, and then walking away until the barrier has worn off (while blink-microing weakened stalkers). Repeat. This is a detail of increasing the efficiency of trading stalkers with immortals that humans don't often even think about, let alone execute (because good blink control is often more impactful). That AlphaStar came up with this shows that it's not just about perfect execution of micro, but also perfect understanding of micro.
There was a "perfect zergling micro vs siege tanks" bot some time ago that would micro lings away from the one that was being fired at by the tanks, thereby negating all the splash damage. The effect was insanely powerful.
But as you say, showing that a bot can have perfect micro is not very interesting. Of course a computer can have better control of well defined tasks like moving a unit away just before it dies, especially doing so for many different units concurrently. What is interesting is the wider strategy and how the computer deals with imperfect information.
The interesting part to me is that, as far as I understand, the AI figured out this strategy by itself, basically deciding that it would be a good way for it to win games, rather than being specifically programmed to do it. That's actually pretty cool!
Other than that, I agree, and am also much more interested in what happens when you have a more level playing field (using camera movement rather than API, limiting reaction times and CPM, etc). I look forward to future matches where this happens.
I think there is some debate about what the neural net did and what was hardcoded. So far all starcraft AIs consist of hardcoded intelligent micro ruled by a neural net that picks one out of less than 100 possible hardcoded choices. And things like "expand", "scout", "group units", "micro" are hardcoded outside of the neural net, part of the API in fact. When the researches said they only used 15 TPUs for 14 days on LSTM, this makes me think they really narrowed down the search space of the neural net and hardcoded a lot of the micro or at least trained separate micro nets.
Not really. The version which learned from scratch was scrapped as it didn’t work at all. This version learned by observing pros. So it didn’t learn by itself, it imitated and perfected pro players.
It was not programmed to do the thing, but all these tactics were in seed replays, from which the agent started its learning. So, it actually not figured the move _by itself_, only found it useful.
I'm curious, would the AI be able to see cloaked units? In sc1 you could see them,( I think sc2 is the same) but it was very difficult. How does the 'raw' interface expose that subtlety?
This is actually a great question. Like what does it mean for a unit to be cloaked?
If humans can, under ideal circumstances, see cloaked units... Maybe the only mechanic that shows up (like for bots or an API) is the inability to be targeted using an attack command (i.e. you can still be hit with splash damage from ground targeting)
My understanding is that the AI sees things via an API the game exposes, so presumably cloaked units are completely invisible to it until they're revealed.
yeah I was disappointed to discover it worked this way.
don't get me wrong, it's a major accomplishment in AI regardless, but it's a significant advantage and it would be easier for me to appreciate the AI's skill if I didn't have to keep reminding myself that it can see the whole map at once. it's such an information advantage.
Actually, I would say this might be the strength of AI from another perspective: the ability to observe and monitor global information without losing attention. Or in other words, attend to the whole picture from get go without being overwhelmed.
While it is an unfair advantage in competitive gaming, but in more realistic settings, there is no requirement that AI needs to have only 2 eyes. It can have as many as it could handle, while human can't scale the same way.
While that would be amazing if true, I'm pretty sure if you take away the stalker blink micro AlphaStar loses hands down to humans. This isn't taking away from Deepmind's victory at all, but I think micro was what made the AI come out ahead in this one. In many of the games, Mana had much better macro only to lose to blink stalkers.
You play the game as it's written. Come back with another version of StarCraft that isn't so micro-intensive and we can see how the AI does on that.
Chess and Go don't have any form of micro and AIs are nevertheless dominant there.
I'd say, give AI development another year and I wouldn't expect there to be any kind of game, in any genre, that humans can beat AIs at. Whether it's Chess, Go, other classical board games, Civilization, MOBAs, RTSes, FPSs, etc.
> Chess and Go don't have any form of micro and AIs are nevertheless dominant there.
Yes, but chess and go have a tiny problem space compared to something like Starcraft. People want to see an AI win because it’s smart, not because it’s a computer capable of things impossible for humans. If the goal was perfect micro they could write computer programs to do that 10 years ago.
Then maybe we need a better game than StarCraft to test this on? Some kind of RTS that's less micro-heavy, perhaps? Maybe even an RTS where you can't give orders to individual units at all, like the Total War series? You can't fault the AI for winning at the game because of the way the game itself works.
Even if you limit the AI to max human APM, it's still going to dominate in these micro-heavy battles because it's going to make every one of its actions count.
> Even if you limit the AI to max human APM, it's still going to dominate in these micro-heavy battles because it's going to make every one of its actions count.
right, and we saw that with the incredible precision with stalker blink micro. There are many ways you could make it more comparable to humans. They have already tried that by even giving it an APM.
> You can't fault the AI for winning at the game because of the way the game itself works.
But it does make the victory feel hollow when it wins using a "skill" that is unrelated to AI (having crazy high APM with perfect precision because its a computer). Micro-bots have been around for decades, and they are really good. The whole point of this exercise is to build better AI, not prove that computers are faster then humans.
It would like if they wanted robots to try and beat humans at soccer, and the robots won because they shoot the ball out of a cannon at 1000 KPH. They win, but not really by having the skills that we are trying to develop.
I just can't help but feel that nothing AI does will ever be good enough according to this mindset, i.e. true "intelligence" is by definition things that computers cannot do.
Beating the world champion in Chess was, at one point, considered an impossible achievement for computers. Now it's considered so routine it doesn't even count as AI according to many. And in a few months when AlphaStar is beating top human players without having to use APM or viewport advantages, what will the next goalposts be?
The point is, it's like being impressed by a calculator because it can multiply two massive numbers faster than we can... no shit, that's the whole reason we use computers, because they calculate faster than we can...
There's nothing impressive in coding something that can execute something far faster than a human, or be so accurate and beat a human. There were Quake 3 bots that could wreck any human alive 10 years ago because they react in milliseconds and shoot you in the head perfectly. So what? It's obvious a computer can do that. It's like being surprised that a bullet beats a human in a fight, that's by design.
I would be impressed if a computer learned from scratch without knowing anything about the game beforehand, about the controls, or anything else, with ordinary human limitations. Using vision processors to look at a screen to see the inputs and controlling a physical mouse and keyboard. That would be impressive. But watching a computer do perfect blink micro at 1500apm is just underwhelming, since that isn't new tech, you could hand code that without deep nets.
> The point is, it's like being impressed by a calculator because it can multiply two massive numbers faster than we can
Yeah, exactly. And when calculators first came out, people were very impressed by them. They upended entire industries and made new things possible that had simply never been possible before with manual calculation. When you're pooh-poohing the entire computational revolution you might want to take a step back and reconsider your viewpoint. It only seems not impressive now because we were born in a world where electronic calculation is commonplace and thus taken for granted.
If you don't find this achievement impressive, then go look at some turn-based game where reaction time is eliminated entirely that computers still dominate at, like Chess or Go. The AIs are coming. Or give it a few months and they'll come back with a version hard-limited to half the APM of the human players and it'll still dominate. It's clear which way the winds are blowing on this. People who bet against the continued progress of game-playing AIs invariably lose.
> Or give it a few months and they'll come back with a version hard-limited to half the APM of the human players and it'll still dominate.
And this is exactly what is being argued here. Let's see that in particular, not a demonstration that computers are faster than humans. Of course they are. Whoever argued that, ever? This has been known and envisioned even before calculators were invented.
What people here are arguing with you for is that we want human-level limitations of the controls for the AI so it can clearly win by better strategy.
> I just can't help but feel that nothing AI does will ever be good enough
It can be good enough in a certain problem space, such as chess. But unlike chess or go, which are purely mental games, Starcraft has large physical component (vision, APM, reaction time). It can make it hard to determine when it has “mastered” this RTS. Like you said, it may be a few more months (years?) before AlphaStar can master Starcraft on “mental” level. The physical component is trivial for a computer, so mastering that is not much of a milestone.
Depending on how you define Chess, seeing the pieces and physically moving them is part of it as well. Chess-playing AIs haven't been required to have robot components because that's not the interesting part of the challenge of Chess. I'd argue the same is true of StarCraft, even more so, given that it's an innately computer-based game in a way that Chess is not. It seems arbitrary to require the presence of an electronic-to-physical bridge in the form of a robot only to then operate physical-to-electronic bridges in the form of a keyboard and mouse. Just let it run via the input devices directly. Give it some years and humans will be able to do this too.
In other words, this isn't an interesting handicap to apply.
> It seems arbitrary to require the presence of an electronic-to-physical bridge in the form of a robot only to then operate physical-to-electronic bridges in the form of a keyboard and mouse.
It's not at all arbitrary. SC2 match is won by a combination of reflexes and physical quickness with which the actions are executed, and strategy.
The whole point is to even the playing field in the area of the physical limitations so that only the strategy part is the difference. You know, the "Artificial INTELLIGENCE" part?
Is a AI that wins at Starcraft only because it has crazy high APM really going to help get to the next X? We could have built that 10 years ago. All it proves is that computers have faster reflexes then humans. That won’t help them become problem solvers for the future.
You seem to forget the way it learned to play every part of the game (not just micro fights). That is, not by having any developer code any rules, but simply by "looking" and "playing".
That's the great accomplishment and nothing like that could have been done 10 years ago.
What makes this interesting is if they can make a computer program better at Starcraft strategy then a human. How they did that is irrelevant. If having developers code rules makes a better AI then deep learning, then the former is the most impressive solution. What they did is a great accomplishment and the AI they created was amazing, but I feel like the faster-then-humanly-possible micro makes any accomplishment hollow, because that is really nothing new.
If they beat human performance in this (non-AI-building) field by humans painstakingly coding rules for specific situations, then that's cool I guess but not groundbreaking, because the solution doesn't generalise.
If they beat human performance in a field heretofore intractable by software by throwing the basic rules and a ton of compute at an algorithm and then waiting for six weeks while the algorithm figures the rest out by itself, then that absolutely is qualitatively different.
The reason being, of course, that if they can find an algorithm that works like this across a wide enough problem space then eventually they'll find an algorithm which will work on the question of "build a better algorithm." After which, as we know, all bets are off.
If you think the how is irrelevant you are completely missing the point of this exercise. Maybe to you only the result matters but for every other task and humanity the how matters.
Simply imagine next taking on a different Game like one version of the Anno series.
If developers did it by hand, you need 50 devs sitting there for probably a couple of months, figuring out the best, rules their sequence and putting them in. That is about $20 Million just to get a similar AI for the next game.
Compare that to download all available replays, requiring maybe 2-3 data scientist to get the data into shape, renting some compute in the google cloud and you get the same or a better result for probably half a million $.
Watch and learn from data alone is why modern machine learning is considered a revolution and novelty. Buying compute time in the cloud is in comparison (to devs and hand coding) dirt cheap and the results are often better.
Deepmind is not working on this problem for the benefit of gamers or the Starcraft community. Making the perfect bot is not the aim. Tackling the next hurdle, next hardest problem in machine learning is. On the way to become better at generalizing the learning algorithms.
Speed of play is a fundamentally important gameplay mechanic of any real-time game. One of the main reasons the pros are better than amateurs at these types of game is because they play and react faster.
And yes, of course computers are much better at doing things more quickly than humans. It's not even remotely close for us. The AIs are clearly better. It's not cheating either; they are legitimately better at it than us.
It sounds like you're simply objecting to pitting people up against computers in real-time games entirely.
So all they really proved is computers are faster then humans. I knew that before this started.
The Deepmind team knows the challenge isn’t to beat humans at Starcraft. That is trivially easy with the advantages you mentioned. The challenge is to be better at strategy then a human. That is why they tried to add artificial rules to make the AI have similar physical limitations to a human (emulated mouse, rate limited actions, emulated screen and visibility). There have been micro AI bots for years that could out preform any human. They knew they weren’t just trying to build another micro bot, because if they were it wouldn’t be much of an accomplishment.
> The Deepmind team knows the challenge isn’t to beat humans at Starcraft. That is trivially easy with the advantages you mentioned.
It's not trivially easy at all. No one had come close before. It took an entire team of ML experts at Google to pull it off. These hard-coded micro bots you're referring to didn't holistically play the entire game and win at it. They're more akin to an aimbot in FPSes, not a self-learning general game-playing AI.
This is yet another in a long string of impressive AI achievements being minimized through moving the goalposts. It's facile and it's boring.
>It's not cheating either; they are legitimately better at it than us.
This is not 100% true, the AI still skips the mechanical part (it doesn't have a mouse, keyboard and hands) in this particular case. This alone can introduce insane amounts of additional complexity, and will make AI to not be pixel precise.
yup. you could have 200 apm, but as long as your clicks and button presses are perfect, you are going to win against someone with 800 but is super imprecise.
blink stalkers are basically perfect for an AI because of the precision they can blink them around.
I assume you’re joking, but just in case you aren’t, Scrabble bots have outperformed top humans for 20 years with little more than a basic Monte Carlo tree search.
In the TLO matchup, the ai wins with an army of disruptors, and unupgraded stalkers; ofc, TLO wasnt playing his best (in terms of micro or race), but it was still doing well with a micro-lacking unit (outside of blowing up its own army repeatedly)
You'll likely be happy to hear that this has been (is being) addressed.
I watched the live broadcast of this announcement where they did a recap of all 10 previous matches (against TLO and Mana) and they talked about this concern. During today's announcement they presented a new model that could not see the whole map and had to use the camera movement to focus properly. The deepmind team said it took somewhat longer to train but they were able to achieve the same levels of performance according to their metrics and play-testing against previous version.
However...
They did a live match vs LiquidMana (6th match against Mana) against the latest version (with camera movement) and LiquidMana won! LiquidMana was able to repeatedly do hit-and-run immortal drop harassment in AlphaStar's base, forcing it to bring troops back to defend its base, causing it to fall behind in production and supply over time and ultimately lose a major battle.
It sounds to me like, although it could see the whole map at once, the fog of war was still applied. So the bot really just got as much information as the minimap would normally give a human player.
> it could observe the attributes of its own and its opponent’s _visible units_ on the map directly
No, not true. Just had an extended argument with a friend over this. Here are some of my arguments against what you're saying:
1. While it's true that a human player could see everything the AI is seeing, the human player has to spend time and clicks to go see those things, whereas the AI sees it all simultaneously without having to use any actions or spend any time to take it in.
2. Emphasis on the computer seeing it all simultaneously. The computer can see the state of two health bars on opposite sides of the map at the same time, or 100 healthbars in a hundred places at a time. A human cannot do that, and even trying to move the view around fast enough to do so would render it impossible to actually do anything else.
3. If it's true that seeing more at once is not advantageous, then it must also be true that seeing less at once is not disadvantageous. So by that reasoning a player playing on a 1 inch x 1 inch screen would not have any disadvantage, since after all they're getting just the same amount of information as long as they move the screen around enough! Reducto ad absurdum, a player with a 1 pixel x 1 pixel screen has no disadvantage either, because they have access to the same information as long as they move around quick enough. It quickly becomes evident that smaller screens inhibit your knowledge of the game state, and therefore larger screen benefit your knowledge of the game state.
One thing they said early on in the ~2 hour video I was watching was that, while AlphaStar had access to the full data of everything within its fog of war, it seemed to need to partition its access to it, in a way that was similar to a human checking different screens, and did so about ~30 (or was it 37?) times per minute.
This might be why changing to having to observe only one screenful at a time (rather than the zoomed out view) didn't seem to have as large an effect.
This is why a lot of competitive games have rightly decided not to support ultrawide monitors. Being able to observe more of the game map simultaneously is a huge advantage. The only fair way to support them would be to cripple the player, by cutting off the top and bottom of the viewable range, not by extending the left and right range.
> whereas the AI sees it all simultaneously without having to use any actions or spend any time to take it in.
Starcraft is a single-threaded game, so I would think that the AI ultimately still has to enumerate through each visible unit one-by-one to collect their information. Why is that so much different than enumerating through each visible screen and then enumerating through each unit on that screen? Either way, the AI could do it much faster than a human, whether it had to click through the screens manually or not. How would it be possible to eliminate this advantage? It seems to me that it's just part of the nature of AI.
No, that doesn't eliminate the advantage -- that's what I'm trying to say. Even if you make the AI move the screen around manually and only let it enumerate units that are on-screen, that's still going to take roughly as long as just enumerating through all the units on the map in one go. It's just a matter of executing "foreach all_units" versus "foreach screens { foreach units_on_screen }". In either case a computer could do that much faster than a human.
Let me put it the opposite way: If you gave the human player a real time list of every visible unit on the map and all of their information, such that they didn't have to move the screen around manually and could see everything at a glance just like AlphaStar can, would that take the advantage away from AlphaStar? No, it wouldn't because AlphaStar could still go through all that data much faster than any human ever could -- no matter how it's formatted or what you have to do to access it. To AlphaStar, checking all the visible screens is just as much work as scrolling through a list of units.
I get what your saying. But screen movement is rate limited (meaning you can't loop through all possible screen positions in 1ms) so you have to actively choose where you want to focus, just like a human player. Think of it more like calls to a web server then "foreach screens".
Can't you click on the minimap to move the camera instantly anywhere on the map?
EDIT: I guess you would still have to wait for the next frame to get rendered, which could add up. True, that does change things a bit, but of course a computer could still do that way faster than a human.
This sounds like a real advantage in the AI's favor though: It can focus its attention on a lot more things simultaneously. It's not just a UI difference; the AI is actually better at this, like how a pocket calculator is actually better at division than people. This latter bit we just accept; we don't defend humans by saying the calculator is cheating because it isn't writing out the calculation by hand.
Similarly, robots are physically stronger than people at any given task you can think of. That's a real advantage of them.
It is certainly a real advantage, but I think the argument is that it's not as interesting as an AI that could win on the strength of better decision-making, or the innovation of novel strategies, etc.
AI wins on the strength of better decision-making and novel strategies in Chess and Go, though. I have no doubt we'll see this in RTSes in the near future as well. For now we may not be quite there yet, as this is simply the first time it's beaten a pro player in any way. Compare with the AlphaGo match vs Fan Hui. A year later and it was dominant over all pro players.
> AI wins on the strength of better decision-making and novel strategies in Chess and Go, though. I have no doubt we'll see this in RTSes in the near future as well.
Yes, likely! I wasn't doubting it's possible or even likely. Only that seeing an AI do flawless 1000 APM stalker micro and macroing perfectly, while pretty cool, is not as exciting as seeing an AI use a novel strategy (edit: especially one that a human could theoretically execute)
I'm guessing that while there's a delay for decisionmaking, there's no delay between when it decides to move a camera somewhere else and when it does move the camera (direct API access), whereas humans need to move the mouse or hit a key, which is gonna take at least like 50-100ms where they're not doing anything else.
When they were talking about delay they were talking about delay between new information -> deciding/acting, which I think obscures the fact that humans have to do new information -> deciding -> acting, where acting takes non-zero time.
{{Delivered in the voice of the female British lady who would narrate Beyond 2000 series of shows - or the Modern Marvels narrator, to your mental predilection}}
After just decades in development, it is clear that the endeavors of those research scientist have finally bore fruit. And today its in the form of:
Intent based modeling, augmented with AI, which provides the reality we see today in both gaming and weapons systems.
The user, who must be human, is provided a range of inputs based on the desired outcome of the interaction with the systems and the real world.
What results, is truly remarkable.
A human is capable of multi-dimensional abstract thought, in a way that a computer cannot. As such - their intent is wired over to a swarm of objects with the most amazing capabilities.
A user can direct a swarm of either virtual bots or physical drones to accomplish a task. She can also combine the efforts of both physical and virtual for even greater effect.
Here we see a swarm of bots who are thwarted by a physical barrier.
The human driver can then instruct his virtual AI bots to attack the security system of the building to allow his drones to have passage.
But she does this through merely the intent for the portal to be open. The bots do the rest.
All the while the user is updated with real-time information on how the operation is progressing.
So, in the future, you may soon see just such technology applied to your kitchen or living room, where bots will cater to your every waking need - and sometimes your non-waking needs as well.
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 was really curious whether they would attempt moving the camera like a human. Sounds like it's still a work in progress, but very exciting! Even this isn't enough to make it fully like a human player, as I believe it is still getting numerical values for unit properties rather than having to infer them from the pixels on the screen. But it seems possible to fix that, likely at the cost of drastically increasing the training time.
The benefit of using pixels, of course, would be that the agent would become fully general. It would probably immediately work on Command & Conquer, for instance, while the current version would require deep integration with the game engine first. But I think the training time would be impractically long.