[jgon]

Even the 200ms reaction time seemed overly slanted towards the AI. I don't think that is the actual reaction time of top pros, in the matches the AI played the human player would teleport in from complete invisibility and try to use an instant cast spell and the AI would have already teleported out. Yes the theoretically may have been constrained to a 200ms reaction time, but in practice the AI was playing at a superhuman level. Even with that advantage in fights, the human team still demolished the AI. Oh well, lots of things to learn still.

[dx87]

Another advantage was that the AI is just reading the game state through an API, it doesn't have to look on the screen. The game can be difficult to watch from a pro's perspective since they have to constantly click around the map to see what's happening, but the AI has perfect knowledge of everything it is capable of seeing, all without having to physically move a mouse to click on the screen.

[spydum]

If you watch the 11th game where pro player wins, (prior was a 10-0 shutout by Alpha), the AI actually lost because they rebuilt the agent to use the same forced camera perspective as the human - so there is absolute truth to this being a compelling advantage. It was able to micro multiple units in disparate areas by having far better spatial awareness. When they took that advantage away it seemed more even.

[rcthompson]

I don't know if we can absolutely claim that the limited viewport was the deciding factor in the 11th game, but it did seem to me that the Alphastar agent's blink stalker micro was somewhat compromised in that game compared to the seemingly superhuman blink micro in previous games.

[ajuc]

You can see the alphastar perspective of that 11th game here: https://youtu.be/H3MCb4W7-kM?t=5195

It struggles with camera placement like real players :) And uses popular divert-attention tactics, which shows it understand that part of the game - for example when it sends oracles to mineral line at the same time as it attacks in front. Previous versions didn't do that, because they were taught playing vs cheating AI - so no point diverting attention of something that has instant access to any unit on the map :)

It also struggles to defend against adept harras beacuse it has "tunnel vision" - controls its oracle instead of defending probes at home. Mana actually managed his attention budget a lot better (this is a crucial pro-player skill in starcraft - harras is effective because it trades little of your attention for a lot of attention of the enemy, it's a skill that becomes irrelevant when opponent doesn't really have "attention" and can perceive and interact with all units on the map at once like previous version of alphastar).

This one is much more human, and much lower level. In my opinion it lost unfair advantage, so the mistakes in its errormaking are revealed. Previously it never was behind and never had to react to human player strategy - it rarely even scouted because what's the point - it wanted to build mass stalkers anyway.

[rcthompson]

Yeah, that's actually a huge point that I didn't even consider. Regardless of whether the AI itself is playing with a limited viewport, the fact that its opponent has a limited viewport opens up the opportunity to learn attention diversion tactics during the training process, which would otherwise be impossible.

[gowld]

What happens if a human tries to use the API with a custom UI of the human's own choosing? Such a UI might not exist yet, but are there ideas for more efficient UIs that could be built?

[confiscate]

Yes I am curious of this too. What happens if the human has a giant TV screen that can see the whole map at once

Or, what if we slow down the game, so that the human can actually pause the game each second and consider what to do next. That's basically what the computer is allowed to do

[NicoJuicy]

It would massively be faster I think.

Upgrade building 3 comes available when you have enough resources.

A separate tab with insufficient resources gives you an overview with what you need to finish a,b,c.

A red alert appears when an enemy is spotted. You can click nearby units attack or a FSM with the attack strategy.

An finished building automatically will be placed near the town center.

Not working farmers can search for resources.

A wall is suggested by your current buildings, you can set an margin of eg. 20 meters.

The question is, how much programming will the custom UI need ( and how deep) to make it a lot more efficient

[dwighttk]

Stalker unit AI could be microing perfectly for you...

[nurettin]

Giant tv:

Macro-wise, it would be like an unwieldly minimap which already exists so people can get a sense of where the enemy is moving. With a giant screen, information is not focused on a small area, so you are limited to your FOV. Minimap which shows unit strength in terms of armor hp or shields as well as placement would be ideal information.

Micro-wise, it would be like sitting in front of a giant text display looking at a whole book. You still have to focus on a small section to read it.

[maksimum]

> Or, what if we slow down the game, so that the human can actually pause the game each second and consider what to do next. That's basically what the computer is allowed to do

While this would make it more fair, it would just make the micro game more similar to chess or go. I don't think humans would necessarily win in the end.

[feanaro]

That's a good insight and yes, humans would probably be overpowered eventually. However, this is just the consequence of the fact that all games are similar if you remove external limitations such as reaction time (or, alternatively, produce a more efficient "being" which is not as subject to these limitations as some other).

Starcraft is like chess in some sense. The largest fundamental difference is that it isn't a perfect information game.

[setr]

Tbh starcraft and dota shouldn’t really be the test games atm; turn based strategies (or rather, grand strategies) would be the far more appropriate evolution after chess and go, since we’re clearly more interested in AI macro than micro, and too much of its learning process is in trying to push the AI beyond micro-oriented thinking (probably many rounds of the AI tournament are lost simply because one AI found a new micro strategy to abuse)

But ofc, there’s no tbs or grand strategy currently out there with a real tournament scene, so you can’t really count on the devs implementing an AI-API, or even properly balanced / bug-free (far more user-testing goes into sc2/dota2 than say civ, simply by virtue of its playerbase).

[zizee]

Well, are we wanting to test the computers ability to strategize/plan, or their ability to out click humans?

The former is an interesting AI challenge/achievement, the latter is a space in which computers are already known to outperform humans.

[ajuc]

Even allowing players to zoom out would give huge advantages, that's why no matter the screen size you have to play at the same zoom. There was a bug at one point that allowed players to play multiplayer zoomed out and it was forbidden to use it in competetive games.

[waterhouse]

How about having multiple humans control the same faction, so one can focus on building, two on a couple of battle groups, another on scouting, etc.? Then they don't have to context switch nearly so much.

[thefreeman]

They actually have this game mode built in, it's called Archon mode.

[waterhouse]

Aha, nice, thanks. Let's see, two players per side... not a huge number but probably a big step up from one. Looks like people aren't playing it much; some people suggest it's because that requires a partner.

I would like to see a setup akin to that of Ender Wiggin, with one commander overseeing and recommending overall strategy, and, say, five others managing different areas or groups. That seems like the way to get the best human performance, and might be enough to beat the AIs—at least to nullify chunks of their advantage.

[thomasfortes]

Yeah, put an eye tracker in a pro and you'll see that the eyes are constantly changing the focus point, if you can watch the entire scene with the same precision without the need to focus on it you're already at a nice advantage.

As an aside, a few pro gamers prefer to play on windowed mode for exactly this reason.

[bitxbitxbitcoin]

Just supporting this. I remember the uproar when 800x600 resolution was removed as an option in SC2 around 2012[1].

[1] https://eu.battle.net/forums/en/sc2/topic/6201040181

[rootlocus]

> the uproar

I'm not saying you're wrong, but 6 posts with no profanity is hardly "uproar" by blizzard forums standards.

[theflork]

Is the bit about reading the game through an api true? Earlier iterations of this same rl based agent that played Atari games would read just raw pixels not an api.

[sytelus]

Yes, it’s true. A special PySC interface was created for AI. Also, it’s not only that AI doesn’t need to parse information available on much limited screen real estate but also that AI doesn’t have to use controller that have physical constraints. So AI has access to this super human controller and it can decide to click on one screen extreme and then another within 200ms.

[setr]

Any game that is specifically going out of its way to support these ai’s will naturally do it through an api, though I’m only aware of dota2 and sc2 (sc:bw also does, through a community-modified client that serves the api, iirc). For adhoc games, eg atari, pixel-parsing is the natural result, but no one would intentionally set it up like that

[asdfasgasdgasdg]

The game is difficult to watch, but does anyone honestly believe that an AI is going to have a difficult time parsing the scene if it is trained to do so? That to me just seems like a question of resources. We're pretty good at image recognition and segmentation now, and that's without the unlimited amounts of training data one could generate when using a controlled game environment with a limited range of possible animations and effects. This is why I find the prospect of the AI agent having to parse the screen entirely uninteresting.

[kaveh_h]

For real life applications, parsing the ”scene” would have impact as it could only convey imperfect information retained. In the game of starcraft the information is perfect when fog of war have been removed this together with unlimited attention (camera viewport) helps action potential and macro planning. No player is ever going to be able to consider precise strategy on the whole map perfectly in their mind. If deepmind wanted to mimic human limitations perfectly they would have to provide imperfect information for AlphaStar, e.g when providing information of locations of objects sample a random variable from a probability distribution which represent the location imperfectly and making that distribution bigger the longer the attention of the A.I wanders from the object both spatialy and temporal. Of course the usefulness of having these limitations is purely to model maximum theoretical human mental capacity and it’s use case could be to help explore strategies that work for actual humans.

[feanaro]

There is another potential use: given these limitations, an AI might be able to learn to be better strategically, which could translate to an even greater advantage once the limitations were removed later on.

[posterboy]

windowing the focus perhaps, yes, but I'd assume it's the opposite and the focus is applied more freely.

[posterboy]

You talk about a static image, but navigating the camera requires strategy, attention, and adds to the focus. If you take that away, it's just a turbo charged pen-and-paper RPG with a time limit on rounds.

They could train against the API, reinforcing the AI trying to predict the state from vision. But with limited APM it would be pretty difficult for the AI to keep track of everything. And, potentially, it would still not be the same as a human looking at it. I'm not sure whether human attention is a particularly bad example of efficient resource allocation. I'm very biased to think it is still the gold standard. But the fact that deepmind didn't focus on this implies they were not finding it interesting enough, and/or too difficult.

Anyhow, (visual) exploration is a step up from mere image recognition

[thomasfortes]

But on the other hand, an AI that beats humans using brute force in a game where it makes a ton of difference isn't much fair too.

[red75prime]

> using brute force

"Brute force" in AI context is usually reserved for traversal of the entire search space. I think "superhuman micromanagement" is a better term. And before AlphaStar superhuman micro wasn't insurmountable obstacle for human players.

[candiodari]

The funny thing is that once we're talking about the real world, which will come, that incentive actually reverses.

At that point the name of the game will be maximizing the advantage the body/infrastructure provides the AI, not minimizing it.

Weird.

[buchanan]

Yes, since DeepMind chose SC2 for having the right characteristics for mapping to the real world, ie imperfect information and real time response, they should have had at least one run without any speed governors. And maybe another with the CPU limited to some level we might find in an embedded system of near future.

[gowld]

It's the same principle as a baseball player putting extra weights on the bat in practice.

[lrem]

I've recently watched a TED talk explaining how human perception has a lag of about a third of a second. Pro players might be better, but after noticing they also need to take an action.

[neltnerb]

My experience is that to beat 300ms requires there to be no conscious thought in the loop. It has to be muscle memory guided by higher level intent. It's like how the gunslinger waiting to shoot hits first, it's reflex instead of decision.

[wnevets]

Getting sub 200ms on something like this benchmark is fairly easy [1]. While waiting for the color to change is different than processing a game like dota2 or sc2 a 200ms limit isn't too unreasonable to me.

I would love to see these AIs get handicapped even more like a full second and really force them to out think humans.

[1] https://www.humanbenchmark.com/tests/reactiontime

[gsich]

I think OpenAI would have been by lots of humans, but they decided to train it with 5 unlimited, invulnerable couriers. (until the TI showmatches, in which they were beaten easily.)