AlphaStar AI goes 10-1 against human pros in demonstration…

On January 25 2019 17:53 alexanderzero wrote:


The term "brute force" has a specific definition in computer science. Generally speaking, it means to compute all of the possibilities and then select the best one. However, this is not feasible in Starcraft for a number of reasons. It wasn't even feasible with the game of go, and Starcraft is a lot more complex. AlphaGo and AlphaStar are both capable of quickly recognizing which possibilities are not worth exploring, but without having to actually evaluate the outcome of such actions. This is why some people describe their mode of operation as more intuition-based. The AI doesn't have to think about what would happen if it made obviously terrible decisions, like move commanding back and forth in front of the enemy's army.



There is no way to spin this as simple PR. Writing an AI that can play RTS at a human level is a world-first achievement in computer science.

EDIT: What OpenAI did with Dota was a PR stunt. As far as I'm concerned, as long as the AI is playing real Starcraft and not some limited version, then it certainly qualifies as a fair evaluation of its strength.

On January 25 2019 17:49 Jasper_Ty wrote:
In my opinion the truest level playing field is not a crippled bot vs a human playing with keyboard and mouse, but a bot vs a human whose brain is directly hooked to SC2, so that the human also gains the advantages that the AI had. Imagine if you also had access to perfect clicks and stalker micro. It's fun to think about-- I think the AI, at least the AI we have at today, would never beat a human hooked to SC2, ever.

On January 25 2019 18:43 Grumbels wrote:

If a professional player elects for a different strategy for each game of the series, then he is still recognizably the same player, with the same patterns and habits in his control, such that you can predict his reaction to feints, charges, diversions. Furthermore, although a player can deliberate in opening choice, they will tend to fall back to their general style in the mid- and late game.

But suppose you play a game versus AlphaStar, you notice not only some higher level decisions such as a tendency to opt for blink stalker builds, but you also pick up on some habits such as bad scouting, inability to defend against harassment, ineffective wall-offs. And now you're considering your approach for next game. It's obviously perfectly proper for the agent to rotate between strategies in order to avoid being figured out, therefore you realize you can't blindly counter the previous blink stalker build. But you also can't count on its weaknesses in defense, because you're playing a different agent. You don't know if it will have good blink micro, because it's a different agent. etc. In its most extreme form, any type of decision which involves interaction with your opponent will be based on quicksand, because you're playing a completely new opponent every time with no history, no information about it. Whatever this is, it's not standard match conditions. It's more like playing five random ladder games versus barcodes, with the opponents being picked out of time and space, e.g. one will be a 2016 player from Korea, the other a 2018 player from Europe, the other a diamond level player that just does cannon rushes.

In short, it's not predictable, unless all the agents tend to converge to a similar style, which is speculative, but if that's the case then I think that humans could adjust and develop anti-AI strategies, similar to how computers could be defeated in chess for years despite superior calculation ability. Obviously AI's will win in the end, but it's an open question whether that's a week from now or five years from now. We don't even know if Deepmind will stay with the project long enough to thoroughly trash all human opposition. afaik there has been BW AI research for around 10 years without threatening pro gamers.

On January 25 2019 19:30 Jockmcplop wrote:
I don't think there's much you can do with a preexisting game like Starcraft to get a 'fair' game between ai and humans. Maybe intense tweaking of balance and ai ability but even then I don't think it would seem like a real opponent.
You would have to design a game from the ground up with ai in mind I think.

I'm very impressed with the decision making of this ai. Its streets ahead of anything else I've seen.

On January 25 2019 19:51 Snakestyle11 wrote:


I think its just super micro potential units that are broken for AI. I promise with zerg AI would not look nearly that impressive. THey would probably have a hard time droning and making units at the right times. Then you dont have things like warp prism micro or blink micro that can scale like crazy. What are they gonna do, dance their zerglings? They cant jump accross a wall. What are they gonna do, shoot an oracle or void ray or banshee with some roaches?

The AI would probably be forced in a ling bane hydra game, and then it would come down to how good they can micro hydras against AOE.

Zerg is the true race where human intelligence shines. Its about being one step ahead, predicting what the opponent will do, where he will send his warp prism,etc. You have to know when to drone and make units sometimes based just on instincts alone or knowing your opponent or current meta trends.

Protoss is you pick a build, execute it perfectly, and if you do execute it perfectly or very close, you probably win,unless your build order was too coin flippy.

In fact, PVP is by far the easiest matchup for an AI like deepmind to win consistently against humans.


FUN FACT: I wonder if deepmind AI could play billions of matches of all matchups to determine which is the potential best race. If someone is unbiased and would know , its an AI. I bet they already know lol.

This kind if AI playing billions of games against same skill opponent could probably easily conclude if certain units are overpowered in certain matchups. Very interesting stuff when you think about it....

On January 25 2019 20:12 CadoEverto wrote:
It seems that AI is cheating.
For example, 4th game vs Mana.
Looking at the corner of own base with selected 2 adepts.
Than the probe is coming (which won't in vision of AI) and place gate. We can see that probe is selected (blue circle around the probe), but at that time in the bottom of screen 2 adepts are still being selected.

From that point we can figure out that:
AI worked not only with 1 active screen (as it was considered).
AI could do multiply selections and actions in one moment, that human can't do.

Please note that the raw interface agents weren’t using the camera directly. The 10 replays have therefore been post-processed to add heuristic camera movements, such that the target location of each agent action is visible on screen.

On January 25 2019 06:44 Waxangel wrote:

This is a really, really insignificant nit to pick. It's a ~200 MMR difference according to their estimation, which is not worth putting in the TITLE. I'll forgive you though because you're just trying to defend mankind's pride.

On January 25 2019 13:00 counting wrote:


This actually has a technical term in machine learning community called ensemble learning. But I don't think it is that easy to implement as of yet. And for efficiency sake the single agent is actually very different from a group of agents which will absolutely require quite a bit of parallel processing to achieve (it is not as simple as installing more GPU can solve). And indeed these agents choose to represent the group of all agents in the AlphaStar league will be those encounter many different strategies and still win for the most part overall. It actually is a very difficult problem to introduce "novelty" and still able to adapt mid-game. The current system is simply not having any learning capability on the fly (within one game, in machine learning term, it is a system with offline learning, instead of active/online learning which is much much more difficult).

On January 25 2019 19:46 LDaVinci wrote:


Maybe I wasn't clear in my point.
I was saying that people could prepare before a match, if they have access to several games for each agent and if they play against the same agents. Then scounting allows you to know which one it is, adapt your play accordingly and choose a winning strategy.
In the case of Mana and TLO, they didn't know before the game that they would play against 5 different agents and even if, they wouldn't have known the behavior for each agent. They couldn't prepare.

It's as if Maru and Serral were playing only training games for a year against each other without sharing replays. It would be impossible to EU Terrans to predict the play of Serral and for Kr zergs to predict the play of Maru.
Strategy wise, they could completely change from one game to another. Of course, tactically wise (micro, reaction to pressure and so on) they are the same player and would have tendancies that would stay the same and you could probably adapt between the games. In that regard the AI and the 5 agents are different. But still how do you prepare for the reactions of players if you only have 1 year old replays

And even in the case of the 5 AI agents, some patterns could still be recognizable : difficulties to split its army, over reaction to seeing enemy units close to its base, and probably others that I missed because I'm no pro. So it could be possible to kind of adapt.

On January 25 2019 15:39 pvsnp wrote:


Point to the cat, please:



It's almost as though humans and computers are very, very different.

On January 25 2019 21:14 Grumbels wrote:

I don't know anything about AI, but wouldn't it be sufficient to simply have the bots play Bo5's against each other instead of Bo1's during the training phase? Because then they can still learn from what their opponent has been doing in previous games.