Amazing progress in AI happening right now

Jow Forums has largely become consumer electronics: the image board co-staring /v/
>where do i go from here then?
imo you're better off going to /sci/, or a different imageboard altogether

If AI is so good, where's my android waifu at?

Attached: large.jpg (500x703, 268K)

Easy for computers to play because turn-based games tend to have simple discrete mechanics with a small problem space

that's easy for an AI.
>look at current state of game
>calculate all possible future states of the game
>find which one favors you most
>make the move towards it
>end turn

then what's according to Jow Forums the best (multi)task to put AI to test

Dota 2

>starcraft
the absolute state
A.I winter 2.0 eletric bogaloo incoming in < 2 years.

Multi task is easy for AI.
Multi factor is what AI has trouble with.

>amazing progress in AI
>inhuman, impossible micro
>impossible APM spikes
>gets obliterated the first time it doesn't have visual access to the whole map at once

yeah nice """progress""", jewgle

Videlo games!!!

Attached: 1533558590116.gif (413x243, 51K)

he says while posting in a video card thread pretending he's using it for important work

>not realizing Jow Forums has always been nothing but gatekeeping retards that don't care about anything but ganoo lincucks, shitty programming projects that will never amount to anything, consumer electronics that most people don't give a fuck about that will never be more than niche products, and arguing over uamdbro vs intlel and uamdbro vs novidia

Want to discuss ai? Try /sci/ or in this case /v/. This board has way too many brainlets to actually hold a worthwhile discussion.

Are you getting paid to post this?

Ah, so that's why in Civ IV V and VI the AI has to churn for about a minute before it makes a retarded decision by any human metric, managing to lose against handicapped human players almost every time.

Why not post your AI research in /dpt/?

You win internet point by shitting on machine learning on Jow Forums.
Mostly because they don't understand it and their knowledge source are the verge articles on killer robots.

If you put a real AI team, like the ones working on SC2 or Dota2 or Go or whatever, they would be able to create an unbeatable AI very quickly. It's not comparable to the AI that ships with the game

>>AI winning a game that's almost entirely mechanics
>>amazing progress
this. does the AI have limitations like the human body while doing those simulations? or can it place 300 perfect actions in 1min?

Bideoi gameess!!

Attached: 9DA9A9F8943442A4877A8156ECA8C2BD.png (644x800, 14K)

>AI cheats and wins
>/v/ basedsniffers go apeshit

it was a time where Jow Forums were full of /prog/ crossboarder and you could actually hold discussion outside bait posts but this time is over and troll posts are the only ones remaining.

Yes. It is. The games would not sell if the AI was actually good because everyone would lose. The AI in CoD aims around the players head and only ever hits if the player stands perfectly still or takes to long to kill the NPC's. Game devs have intentionally been making stupid AI's since difficulty settings.

Glad you are learning.

>calculate all possible future states of the game
The observable universe contains approximately 10^78 to 10^82 atoms.
Go has approximately 10^170 possible game states.

that's why the bot never goes first.

Do you even play videogames or are you just pretending to be retarded with that example?

>Go
>>>/csg/

are you implying video games are less complex than go

What's retarded about that example?
It's a popular turn-based game and there are many video game implementations of Go.

>AI
>/sci/ or /v/
While there mightn't be much discussion on Jow Forums about AI, and the few threads here are short, at lest it's actually talking about real AI, anywhere else is just philosophy majors seeing who can ejaculate the furthest distance.

You don't need to calculate EVERY possible game state, just the ones relevant to the current position.
With videogames most of the time that's a lot easier then GO.

Most of them, yes.

>You don't need to calculate EVERY possible game state,
explicitly wrote
>calculate all possible future states of the game

so which ones are relevant?

>Most of them, yes.
nope

Go has a 19x19 grid and two possible pieces
Civilization has 100x100 grids or greater with hundreds of possible units onto of dozens of terrain types
And that's just a grid based game
imagine a game where unit positions are 32 bit floats

also most turn-based video games contain hidden information, e.g. fog of war or random number generators
AI so far has only excelled in turn-based games with no hidden information, namely Chess and Go.

Take Civ for example, you have to pick a victory path very early on, then just progress on an almost linear path to that victory objective.
All the AI would have to do is find the most optimal path, then just calculate for enemy interference.

Most videogames add "fun" less optimal choices because from the human perspective it would be extremely boring without that. The AI just has to not pick those choices.

A game like Go does not have anything like that.

tl;dr I'm saying game devs are shit at balance and that's why it would be easier for machine learning to find the best approach.

How about doing something useful?
>automated theorem proving
Hasn't progressed for 20 years or so.
>stock market prediction
Current AIs perform about as well as naive 1/N portfolios.
>image recognition
All currently used methods are similar to linear regression, so there necessarily exist misclassifications.

Lots of chances for breakthroughs, but current ML isn't up to par. Speech recognition is pretty good nowadays, though.

See
It does not matter that Civ has 100x100 grids.
The player can't even see them and can only access around 9 of them at the start.
Venice in civ 5 for example, can only have a single city.
Playing "tall" strategy is just about ignoring everything and building a few really good cities and tall stats are the current meta.

Imagine writing an AI that played the Halo campaign.
All it'd need to do is run VSLAM, some object recognition, and teach it to click on buttons. Little of any intelligence is needed

the shit "ai" has map hacks and can see through the fog of war. yeah totally a breakthrough you braindead rat

I wasnt disputing it would be easy to write a civ AI, but complexity means computational complexity

current AI has toddler level performance at anything but pattern recognition problems
there's very little actually useful coming out of all of this other than facial and speech recognition and almost all cases basic mathematical approaches outperform AI tremendously - and by basic I mean generic regression or statistical stuff most high schoolers could do

>can see through the fog of war
it can't

If I sat down and did the math right now I could \write a simple script right now, that would always win at civ5.
Not because of my amazing programming skills, but because civ5 is not balanced at all, and most games are the same.

Effective APM, which makes it even more impossible.

sure buddy, I know you can do it ;^)

>or can it place 300 perfect actions in 1min?
That what actual pro players do.

The AI actually performed less actions per minute than the pro players it beat.

their actions aren't perfect
most of them are just flat-out redundant

>or can it place 300 perfect actions in 1min?
Yes, it's also 250 on average per minute. (perfect actions, mind you)
It was able to do 1500 (effective) apm in short bursts...

Civ is unbalanced so while it has hundreds of possible units, it really comes down to a few. (archer spam)
Either way, creating a strong AI for it would be hard, yes.

They should focus on Stratego AI first, probably.
It's extremely complex for a computer (imperfect information, big board)
Though I'm sure they could do it if they just throw hardware at it, I hope they can get it optimized enough that it doesn't need 1000W of tensor cores to get it at a strong enough level.

Most of the improvements in "AI" have been hardware more than software.
MCTS algorithms have improved yes. But without Google's TPU (or Nvidia's 2xxx series emulating Tensor cores) the AIs are fairly weak.
Look how LeelaChess0 (open source alpha0, chess engine) performs on GPUs that can't emulate tensor cores or how it performs on CPUs vs how it performs on the latest RTX cards.

So is it amazing progress of "AI" or is amazing progress of hardware?
NN-engines (and genetic algorithms) are as old as the first checkers programs, the difference between now and the past is that we have hardware that can take advantage of more scaleable algorithms.

Also Alphastar had a lot of trouble in its early development cycle. So they had to resort to training the AI on very small limited maps teaching it one thing at a time.

And Stratego has even more...
>possible
And it's a lot of 0s less if we're only talking about ones that are reachable during normal play.

>If I sat down and did the math right now I could \write a simple script right now
I don't think a simple minimax function would be enough to create a strong civ5 AI...

Why do you even post when you know that you have no idea what you're talking about. I never understand posters like you. It's like you're taking a scattershot guess and seeing what hits

The thing with GO, like you said 19x19 grid with two options.

Most video games don't have anything close to that computational complexity. Most games start you with very few options and grow in complexity from that. Because both sides start with the same few options, the AI can just pick the most optimal choice and snowball off that. Let's take starcraft, if the AI is zerg and builds a spawning pool, it does not have to calculate for roaches, or ultralisks, or any other zerg unit besides zerglings, and it does not have to calculate every unit the opponent could have, because at this point in the game, the opponent can only make small selection of units.

From that point complexity will grow, but it won't mater because the AI will already be at an advantage to the opponent, and in SC2 advantage is compounding.

Attached: 250px-Cassandra[1].png (250x269, 69K)

You don't actually know what computational complexity means

starcraft is really not of interesst here, mainly because imposing human limitations on a computer misses the point completely

there are enough turn based imperfect information games around that have huge complexity and require strategies most people would classify as "creative"

You are correct, i was going off a different definition.

the comparison was fucking lame

>the games against TLO, the AI had visibility of entire map minus FOW, meaning no camera movement needed
they fixed this later
>APM high as fuck, since they compared to players, not knowing at least 60% of the APM of real players are useless clicks

the challenge lies not in making a stronk ai, but one that loses in believable ways

What's wrong with APM? It's a machine, of course it's going to react faster.

Attached: 1528487050863.jpg (645x729, 81K)

>I could \write a simple script right now, that would always win at civ5.
Congratulations, you're now a millionaire.
That is what Firaxis will pay for a perfect Civ AI that doesn't cheat.

Because they're trying to create an intelligent AI.
If it can only win thanks to a massive mechanical advantage it isn't really intelligent.

>AI being used primarily for advertising and fucking videogames
its shit, zoom zoom.

>CV
>Linear Regression
pick one fgt

no they wouldnt
it's not an unsolvable problem, it's just a problem they aren't willing to spend the resources on to solve
they dont use machine learning to write their AI, they hand-tweak some tables
good AI isn't that important to video game sales

>loses half of its army to the same rookie mistake consistently
>"mechanical advantage"

AlphaStar averaged less apm, it just made better moves.

NN approach would make it a nightmare to get mods working.

you can release training tools with the SDK if you care about moddability

You're forgetting that players combine mods and it would be players who'd have to spend aeons training the net to work with their particular combination of mods.

It is such a shame they didn't limit it to like 50 APM or something low like that. Most of the human APM is not effective anyway. They should also make it play on ladder so you could take a look at all the cases where it loses. It seems the game with MaNa made it shit itself because of the harassment.

It wouldn't take long to train a Civ AI, and you can train it with specialist strategies that fall back on generalist strategies if you're concerned about mod compatibility
Firaxis really do not invest resources into making AI, no developer does, it doesn't really make sense because it doesn't sell your game

Kinda disappointing desu, I wished the things would run VSLAM to get map data, force some scouts to explore and all that shit.

It's going to be a nightmare to use mods with NN AI and that alone is a reason good enough to not have that in a civ game.

> super duper AI
>it's absolute weakness are cheese builds, troll builds and human creativity dicking around

OOF

>you can train it with specialist strategies that fall back on generalist strategies
I have no idea why you're defending the status quo like this couldn't be done
it absolutely could, and it wouldn't even be difficult

You are underestimating how much computational power you need to train a neural network.

lmao

>cheats
Wow what a surprise it wins.

You're overestimating the complexity of a neural network needed to play Civilization
You could produce meaningful results training one on your own PC

But he said he could do it easily.

If I train it for days. You clearly have no idea how retarded those things are.

>If I train it for days
are you implying that's a long time? For serious applications you need entire clusters of computers. A neural network is not "retarded" or smart, it's a tool with an application. Video games basically use NNs with hand-coded data as it is

>days
You're a moron of the highest order.
If you're trying to do it on just your home-PC you would be talking about years of continuous training.

That is a long time for a person who installed mods, doesn't care about AI and wants to play. 10 minutes is a long time. And that's pretty much everyone playing civ.

Don't fucking reply to me, reply to the guy claiming it's doable.

>If you're trying to do it on just your home-PC you would be talking about years of continuous training.
That depends entirely on the application
Civilization from nothing to godlike ala Dota 2 obviously isn't practical
You would break it down into specialized units like a tactical AI that controls army units or a strategic AI that makes long-term game decisions

And do you have an idea what kind of evaluation function to use?
Best-first or depth first?
MCTS? A/B pruning?
Do you have any experience writing an engine for a board game? Can it beat a weak one like Fairy-Max?

You don't need to have an AI that you re-train based on mods. You can have an AI that acts on generic statistics by examining the properties of a unit or your incoming food/gold/production so it wouldnt need to specifically understand every item to be able to play. It knows you have X amount of incoming gold regardless of whether you have a vanilla bank or a special modded bank

Do you think throwing out ML jargon makes you sound smart? I program these things for a living, I am experienced with video game AI

How does it know how many banks a player has?

Sure you do.
Can you show me an example?

>a/b pruning is ML jargon
lmao, fuck off retard

>averaged
Exactly, user.
Average. AlphaStar has insane APM spikes.
It learned to handle apm as a ressource because the devs limited the apm average.