Starcraft 2 Artificial Intelligence

Project AlphaGo SC2 Deepmind has begun.

youtube.com/watch?v=WEOzide5XFc

youtube.com/watch?v=6L448yg0Sm0

youtube.com/watch?v=-fKUyT14G-8

Any Veeky Forums predictions on if it will beat professionals or display top tier strategy at least?

>Alphago SC2 vs Innovation
>Alphago picks zerg
>6pools every game
>wins

I'm just interested how it would even work. There could be very interesting race vs race vs race matchups with alphago playing itself due to the assymetric nature of the game.

The problem is the micro vs human input aspect. AI has tackled 1v1 poker already so it should be able to handle the "mindgames" portion of 6 pooling vs expanding vs standard play.

It will be much more interesting than Go in terms of watching it but I have a feeling the micro aspect is going to ruin any true human vs machine comparison. Not that it didn't have huge computational search advantages vs humans in go, but the 200ms human response time and physical input is just completely outclassed micro wise by a computer.

Kinda wish they had chosen civilization games.

Honestly speaking, I think Alphago will beat the koreans easy.
There is already built in AI, and there were user map settings (modded maps basically) back in SC1 with upped AI difficulty.
Imagine some non brainlets with the budget try to make that on steroids.

AI will have perfect micro, so the only way a human could actually beat it is with some exploitative strategy.

I'm not investigating because I'm interested in it for entertainment.

I believe they are putting some artificial limits on APM for the computer. They should also restrict it's reaction time and some other things to "human level" even if it has many more cycles of thought to plan and search for solutions.

Sc2 is a kind of game where no human is remotely close to optimum play, so computer is at a great advantage

If its APM/reaction time is limited, then koreans may win because of the sheer amount of unpredictable plays that are possible.

In SC, it is not possible to have 100% information and there will be plays that can't be countered (have no real solution).

Yes, in such games it's not about 100% win rate but about maximizing win rate. Hence even the most skilled players will use unsafe openers just because if they played safe every game it would be a lower win %. It would be interesting to see if alphago develops cheesy openings and how it begins to develop a varied set of opening strategies and hidden tech etc.

intere

This. The ingame AI in sc1 was pretty hard to beat. There's no doubt in my mind even pro players will come even close to beating a dedicated sc2 AI.

I'm not so sure. Alphago might discover some hereto unknown strategy that completely shits on everything before that nobody ever thought of because it's so unintuitive.

The big problem with deep learning is it doesn't think (or anything close to it). You basically just train with a ton of 'images'--some n-dimensional, order-dependent representations, such as game replay data. So there is never any understanding of why something is good or bad or the trade off of pros and cons, or the complex interrelation of concepts, etc.

It just doesn't seem possible to me for pro players to lose with all the variation available in SC2. For reference, AlphaGo can only play on standard sized boards. Every single professional go player can dominate on any sized board, but because all the training data is 17x17, AlphaGo will only get good answers for 17x17, because it doesn't understand any underlying strategy.

>ingame AI in sc1 was pretty hard to beat

kys for posting such idiocy, RTS AI is one of the absolute worst.

It really depends. Maybe with automatic perfect micro there are some. The thing is the early game is the most played by humans and they test all of the boundaries. I'm sure the AI could eventually come up with maybe more optimal starters but there will always be a risk vs reward to different strategies because of the cost of scouting.

It's going to be entertaining but I wish it was broodwar instead of sc2.

Sc1 AI was pretty vicious desu

It's not a problem. It's a pro/con. Everything has them. There are energy usage differences, scalability differences, etc. It's just not comparable along 1 single measure and then saying it is bad.

fuck sc2, do this for the civilization series. those games have horrible AI and since most people play the games in single player, we have to suffer an awful experience.

Jesus christ why not SC1, the opitomy of a game with hidden complexity and an infinite skill ceiling and changing meta.
It would be amazing to see if it managed to go through similar changes in playstyle as the long standing pro scene had, SC2 is just garbage and has no pro scene or history to compare too.

i know nothing of this, but random question. Were Perfect Dark's "perfect sims" beatable? I just want to know because i played that game obsessively and those perfect sims seemed to cheat physics and logic

Deepmind should monetize AI services. $10 a game to have a Civ game against good AIs.

>implying a computer could handle muh mutacontrol

I want to see how an AI reacts to getting cheesed

I would love to see how it develops. Hopefully they document the early games a lot and show off interesting strats it tries.

All it's gonna do is cheese

IIRC, SC1 AI didn't play by the same rules though
like it didn't need Gas to build stuff

this. Cheesing every game will give it the highest prob of winning for most maps. And for larger maps it will go for a 15-min all-in timing attack.

Every non-pro knows in his gut that these tactics give the highest reward, but we don't always do it because we want to learn to "macro".

not really. The travel time of worker makes it inefficient. The player going more defensive will win, especially with wall ins + pulling workers.

whats so novel about it?
sscaitournament
SC1 AI tournaments have been going on autmatically for almost a decade on twitch and it beat the worlds best players

one of the bots has insane mutalisc micro

If it has infinite APM it will have no issue beating pros. If they limit APM and reaction time it might have a hard time depending on how good the researchers are at setting the metrics for success. If the researchers do well, we should see top tier strategies, possibly even interesting strats.

My personal hope is that the AI goes for a "never use the same strategy twice" approach, as then the pros can't train to counter it as it can surely train to counter them

they don't beat the best players

Because the SC1 AI literally cheated and knew where you were at all times, didnt need gas, etc.

This AI is a more 'true' AI that only relies on what it sees onscreen like a human being.

Will it master the artosis pylon?

The issue with 3+ player poker games is that there are counter strategies.

You could play the nash equilibrium that always wins (ties) if everyone else is also playing nash equilibrium or some substandard betting method.

However if another player knows you are playing nash equilibrium they can use that knowledge against you.

Koreans BTFO

>implying this wont make the koreans stronger.

Very true. No one will be impressed with a AI SC2 victory without I/O limiting. All we will see is micro exploitation.