"AlphaGo" Deepmind Starcraft AI announcement

File: 130 KB, 640x360, deepmind-1.jpg [View same] [iqdb] [saucenao] [google]

"AlphaGo" Deepmind Starcraft AI announcement Anonymous Wed Jan 23 20:36:03 2019 No.10321959 [Reply] [Original]

Deepmind Starcraft update to be shown off live soon

The most recent developments will be revealed at 10 am PT/1 pm ET on January 24, on the StarCraft Twitch channel and Deepmind's YouTube channel. Hopefully it won't include anything about viruses.

https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A/featured

>>	Anonymous Wed Jan 23 20:43:48 2019 No.10321974 >>10321959 Oh shit, I was just searching for an update on their progress yesterday. Amazing timing. Thanks anon.

>>	Tuba Jam Wed Jan 23 21:13:06 2019 No.10322037 >>10321959 BUMP this is going to be huge anyone know a video that gives a crash course in the basics of competitive starcraft so I'm not totally lost during the match? also, >inb4 1000 apm from DeepMind

Anonymous Wed Jan 23 21:27:12 2019 No.10322068

>>10322037
I can give the basics. Basically there is an extra resource of attention in RTS games. So for instance you can trade off economy for better military power by focusing more on micro instead of economy.

This means a key consideration is where to look, whether to glance down at the miniature map in the corner to see anything suspicious etc.

The other concept is that you have a choice between enhancement of economy, technology, or of more military strength, with some rock-paper-scissors elements to that, which vary based on attention requirements to utilize the units.

This means there is complex decision making at all times. Broodwar (Starcraft 1) is actually a much better competitive game because the "attention" management is even more important which allows comebacks and long games to be extremely skill dependent.

APM - Actions Per Minute is how many orders you give per minute. This generally is maximized by players as sort of "pace" setter to get their actions done fast. So as to maximize attention rewards.

Anonymous Wed Jan 23 21:28:42 2019 No.10322071

>>10322037
best bet is probably playing a few games to get the basics and then watch a few commented matches
I like lowko https://www.youtube.com/watch?v=0cvE-6KziCE

most of players APM comes from moving the camera around all the time and swapping between production buildings to check progress, not sure whether deepmind is more efficient about this.
if all of those actions were to be spent on unit micro though it would be pretty sick

Anonymous Wed Jan 23 21:29:49 2019 No.10322077

>>10322037
Well, from everything I've read over the last couple of years about deepminds attempts at playing SC have been abysmal.

In fact, just recently, they stated something along the lines of (can't find the quote, as now google is flooded with the announcement in the OP), that a dramatic breakthrough in machine learning was required for neural nets to stand a chance at beating the most rudimentary starcraft 'bot'. This was about a year ago though.

So unless something seriously has changed over the last year, I wouldn't get too excited.

(I can't fucking wait for AI to actually be able to challenge the best SC players, though)

>>	Anonymous Wed Jan 23 21:30:51 2019 No.10322080 >>10322037 Also, APM will be limited to within human constraints. Rudimentary computer bots with unlimited APM can already crush humans, because its essentially cheating, not exactly 'thinking'.

>>	Anonymous Wed Jan 23 21:31:47 2019 No.10322082 >>10321959 Brood War or Starcraft II Brood War would be much more impressive considering there's a lot of janky shit you have to do to compensate for the broken pathfinding.

Anonymous Wed Jan 23 21:37:18 2019 No.10322093

>>10322082
SC2.
>>10322077
https://news.blizzard.com/en-us/starcraft2/21509421/checking-in-with-the-deepmind-starcraft-ii-team

As of March 2018, Deepmind was complete shit at SC, and the engineers said 'they didn't know' if it would ever compete with humans.

>>	Anonymous Wed Jan 23 21:41:04 2019 No.10322099 >>10322082 Broodwar would have been the better choice. I hope they eventually try the alphazero approach of getting it to work on all big RTS games.

>>	Anonymous Wed Jan 23 21:44:15 2019 No.10322108 Fuck. Hosted by Artosis and Rotterdam. Where the fuck is Tasteless? Get this rotterfaggot the fuck out of here.

Anonymous Wed Jan 23 21:45:40 2019 No.10322111

>>10321959
It's hard to compare an AI to a human in a RTS game.
The AI will have a speed advantage in reaction time and will be a god at micro-management.
The speed advantage could allow it to dominate even if its strategy is shit.
It would be cool if they limited its speed to that of a human.

There would also be an issue regarding map-specific strategies.
If they only trained it on one map, it might just optimize for that map.
If presented with a new map, it might play like shit.

>>	Anonymous Wed Jan 23 21:46:17 2019 No.10322114 >>10322111 AI is limited to physical human constraints.

>>	Anonymous Wed Jan 23 21:48:41 2019 No.10322126 >>10322108 Rotter shills SC2 a fuck ton. It's also blizzard in charge and they are an absolutely dreadful cancerous company right now.

>>	Anonymous Wed Jan 23 21:48:54 2019 No.10322127 >>10322077 Can somebody explain what is the difference between AlphaGo and the usual AI in games?

>>	Anonymous Wed Jan 23 21:49:17 2019 No.10322130 >>10322114 Is it limited in just speed or did they also limit it in mouse precision? If it is making perfectly accurate clicks, it might have an aimbot advantage.

Anonymous Wed Jan 23 21:49:54 2019 No.10322132

>>10322111
>It's hard to compare an AI to a human in a RTS game.
Yes, it's less interesting if it can maximally use a micro track vs if it has interesting strategies

>>10322114
don't ever believe this. The constraints would be enormous to do this, affecting it's vision time, reaction time, how many clicks it can do in a very short time period etc.

Anonymous Wed Jan 23 21:51:40 2019 No.10322138

>>10322132
>don't ever believe this. The constraints would be enormous to do this, affecting it's vision time, reaction time, how many clicks it can do in a very short time period etc.
You believe they can create an AI to outperform pro players, but you don't think they can impose limitations on how it can play?

Anonymous Wed Jan 23 21:54:28 2019 No.10322146

>>10322138
I simply doubt that deepmind or blizzard will care all that much about anything other than an APM limit.

It's not that they can't do it. It's that they won't because saying "it's max APM is 350 just like a human" is enough for the media and public. Plus actually giving it equivalent mechanical constraints would be a lot more work on both blizzard's end and deepmind's end

There is no incentive to do so.

>>	Anonymous Wed Jan 23 21:55:40 2019 No.10322149 >>10322146 >There is no incentive to do so. Except, you know, actually test their hypothesis.

>>	Anonymous Wed Jan 23 22:05:53 2019 No.10322173 I want to play against deep mind in coop. Instead of spawning waves to send at us they should make deepmind control bases and produce units to send at us.

>>	Anonymous Wed Jan 23 22:06:55 2019 No.10322177 >train robots in military strategy so that humans will lose every war against them Idiots at blizzard surprise me every time

>>	Anonymous Wed Jan 23 22:29:14 2019 No.10322227 >>10322177 go to bed yudkowsky

Anonymous Wed Jan 23 22:31:56 2019 No.10322236

>>10322037
You got to build shit efficiently, in the right order and timings, scout and control units to defend or attack.

Honestly, I'm not surprised they cracked it since they already solved DOTA which proves their AI can learn micro and counter oponents' builds. Get a machine learning algo running on the opening/macro theory and I believe it can be done. It is left to know if it plays better than humans and is able to figure out new ways of playing that are not in the meta, like fancy cheeses and shit like that but I think it's possible

>>	Anonymous Wed Jan 23 22:34:27 2019 No.10322241 Inb4 they brute forced using plain vanilla PPO and 3 billion machines like in DOTA

>>	Anonymous Wed Jan 23 22:38:22 2019 No.10322257 >>10322077 Me here. I take back what I said. Here is the latest from deepmind, Novmber 2018 at Blizzcon. Get ready. Deepmind is about to crush the koreans. https://www.youtube.com/watch?v=IzUA8n_fczU&t=1361s

Anonymous Wed Jan 23 22:57:44 2019 No.10322323

>>10322149
you're a fucking retard, seriously fuck off low IQ piece of sewage filth

Do you think they care? How often will it glance at minimap? How often can it read an entire screen in a millisecond, is it going to actually be exactly constrained to human mechanical and visual capabilities? Do you think they will have it aware of mineral count at all times or have to "glance" up via some physical simulation constraint?

You're literal human garbage if you are stipulating it is obvious they will actually give a fuck about such exact competition. Not to mention inherent incentive of both blizzard and alphago to have an AI that competes well

seriously shut the fuck up, stop posting, fucking LOW IQ MONKEY. Your fucking shit 100 IQ brain should not fucking conversate online.

>>	Anonymous Wed Jan 23 23:01:14 2019 No.10322336 >>10322323 Hehe, you are literally autistic, aren't you? This is all additionally funny because of how wrong you are, on basically every account.

>>	Anonymous Wed Jan 23 23:01:32 2019 No.10322338 >>10322149 If I wanted a post like yours, I would go to the bathroom, take a huge shit, smear it on my monitor, and read it. THANKS for saving me all that work to be equally disgusted

>>	Anonymous Wed Jan 23 23:03:56 2019 No.10322345 >>10322338 Hello again faggot, please refer to >>10322336

>>	Anonymous Wed Jan 23 23:04:17 2019 No.10322346 >>10321959 Starcraft?

>>	Anonymous Wed Jan 23 23:45:56 2019 No.10322451 >>10322345 wow dude that source of how they are equalizing capabilities was great >300 apm cap is very different than equal capability.

>>	Anonymous Wed Jan 23 23:46:49 2019 No.10322452 >>10322451 180* restricted to physical camera view* can only see what human can see*

Anonymous Wed Jan 23 23:48:43 2019 No.10322455

>>10322452
so it can read minimap every single frame
so it can read minerals, gas, and population every single frame
so it can switch views at superhuman speed and read each screen that would be a blur to a human

great job fuckhead, seriously stop posting, too low IQ

>>	Anonymous Wed Jan 23 23:49:47 2019 No.10322458 >>10322455 No... it has built in latency you fucking subhuman. Go do your own fucking research.

>>	Anonymous Wed Jan 23 23:49:49 2019 No.10322459 AI brainlet here. Are these advancement actually impactful or dangerous, or are they a series of one trick ponies that don’t generalize well?

Anonymous Wed Jan 23 23:50:42 2019 No.10322462

>>10322177
starcraft has almost zero connection with military strategy brainlet, the people at CENTCOM and the DoD aren't training their analysts with starcraft. Its a basically automatic self organizing system backed by massive mil-industrial complex and a world rigged with thermonuclear weapons pointed at every conceivable target. The subhuman spergs wasting their frontal lobes on this gay shit couldn't strategize their way out of a kosovo tier conflict much less a full scale nuclear war.

Anonymous Wed Jan 23 23:53:25 2019 No.10322470

>>10322458
>No... it has built in latency you fucking subhuman. Go do your own fucking research.

ROFL
>it has built in latency
means literally nothing and does not mean equal capabilities if it can read the minimap every single frame , precisely click exact units it wants, etc.

>>	Anonymous Wed Jan 23 23:55:07 2019 No.10322473 >>10322470 >ROFL How old are you? You know you need to be 18 to post on these boards, right? There is built in latency for every process it makes, minimizing any 'super human' advantage.

>>	Anonymous Thu Jan 24 00:21:36 2019 No.10322528 >>10322473 >minimizing kek, new definition of equal

Anonymous Thu Jan 24 00:23:12 2019 No.10322532

>>10322473
>equivalent mechanical constraints from >>10322146

is the basis of the argument

saying "minimizing" is changing the argument. You have already conceded you are shit for brains and wrong.

They did a few simple things but are not creating equivalent constraints.

>>	Anonymous Thu Jan 24 00:23:18 2019 No.10322534 File: 127 KB, 645x729, b90.png [View same] [iqdb] [saucenao] [google] >>10322528 >huuuurrr the computer program isn't exactly equal to a human meat bag the whole experiment is pointless

>>	Anonymous Thu Jan 24 00:24:41 2019 No.10322538 >>10322532 >equivalent constraints. Except this never was my argument, that was yours. Please return once you have basic reading comprehension.

Anonymous Thu Jan 24 00:28:53 2019 No.10322554
File: 38 KB, 556x556, 1094504fc7a5dd4f4a293590d19c2d86.jpg [View same] [iqdb] [saucenao] [google]

is this an OK formulation of their hypothesis, that the anon higher up is sperging out about?:
>ai plays however it can at standard playback speed, not allowed extra processing time while humans are thinking/playing
>human is allowed to play at any gamespeed desirable (e.g 0.1x speed), potentially archon-mode with team of humans (several humans can control the same players units), pausing anytime
>... --> can the ai compete with the humans in terms of strategy?

nevermind the viability of a human spending 20 hours on a single game

Anonymous Thu Jan 24 00:30:31 2019 No.10322557

>>10322554
>>human is allowed to play at any gamespeed desirable (e.g 0.1x speed), potentially archon-mode with team of humans (several humans can control the same players units), pausing anytime
I would actually really like to see two Korean gaming teams play like this.

>>	Anonymous Thu Jan 24 00:33:54 2019 No.10322566 >>10322557 not 0.1x, but... https://www.youtube.com/watch?v=5YwZRUjF8JY

>>	Anonymous Thu Jan 24 00:34:39 2019 No.10322569 >>10322566 Cool, thanks.

Anonymous Thu Jan 24 00:37:40 2019 No.10322580

>>10322534
>there is no incentive to (give it equivalent mechanical constraints) do so

>>10322149
your reply.. is saying that they would have to give it equivalent constraints to test their hypothesis.

think you need to find a bridge and see if you can fly

>>	Anonymous Thu Jan 24 00:37:45 2019 No.10322581 >>10322566 I'd like to see a game more played like chess. You have two teams on either side, but controlled by one player. The game is unpaused for a certain period of time, say 10 seconds, then paused again, discussed, then unpaused again.

Anonymous Thu Jan 24 00:38:41 2019 No.10322584
File: 56 KB, 471x614, building clearing.jpg [View same] [iqdb] [saucenao] [google]

>>10322554
shouldn't fucking matter. Starcraft's a (continuous?) partially observable markov decision processs(POMDP). Just showing that we can have a computer beat a human at a POMDP type problem at any speed is a pretty big deal. I mean POMDPs are very important, especially this one, which totally doesn't have any immediate application to other continuous POMDPs with military impact. It's not like POMDP solvers have been used to solve problems the military faces like optimally neutralizing rooms or anything...

Anonymous Thu Jan 24 00:40:43 2019 No.10322590

>>10322580
The incentive is to minimise its super human capability, thereby prooving they have actually created something that can 'out think' a human, and not simply outplay it with super-human capabilities.

Might I suggest finding the nearest rope and hanging yourself.

Anonymous Thu Jan 24 00:44:38 2019 No.10322598

>>10322590
Name a human that can watch the minimap, know it's mineral count, and see everything on the screen clearly at one time

Name someone that can click perfectly on each unit.

Again, the gist of the argument, was literally that they would not make it equivalent capability to a human and would leave in advantages.

You have only moved the goalposts, which was exactly what I stated to begin with, that would do some APM and half-measures but had no incentive to actually care about equal capability and would retain advantages.

>>	Anonymous Thu Jan 24 00:45:41 2019 No.10322600 >>10322590 now your retard brain has changed it again from the initial position of equal to "good enough" but retaining advantages which is exactly what I pointed out that you argued with. Shit-for-brains scum

>>	Anonymous Thu Jan 24 00:48:41 2019 No.10322610 >>10322600 you sound very insecure

Anonymous Thu Jan 24 00:51:26 2019 No.10322619

>>10322610
lel

>Hey guys I'm sure it won't be equal, they will do something like APM but won't have incentive to actually ensure every mechanical capability is equal

>NO THEY WILL MAKE IT EQUAL BECAUSE "ACTUALLY TEST HYPOTHESIS"

.... retard shitbrain continues argument

4 posts later

>yeah it wont actually be equal and now I will argue something else

low IQ

>>	Anonymous Thu Jan 24 00:56:20 2019 No.10322627 File: 409 KB, 750x1111, thrad.jpg [View same] [iqdb] [saucenao] [google] you guise are so hostile to eachother

Anonymous Thu Jan 24 01:00:44 2019 No.10322640

>>10322627
Yes, because I had already calculated everything related to my post, the incentives of each party, and what level of "equality of mechanics" they would go for. His layer of thinking way like the 5% point on my analysis and I knew it to be wrong because I had already emulated his thought process.

>>	Anonymous Thu Jan 24 01:08:11 2019 No.10322658 >>10322640 >His layer of thinking way like the 5% point You're embarrassing yourself.

>>	Anonymous Thu Jan 24 01:12:12 2019 No.10322671 Let me know when it can play cpma/warsow

Anonymous Thu Jan 24 01:18:56 2019 No.10322690

>>10322658
I proved my point, read thread, done.

low IQ people and public like who I replied to will be satisfied they are equal if they put in some limitations and say "it's equal" regardless if AI still has huge mechanical advantages

I literally explained this in the thread. The Incentives don't exist because low IQ people, AKA person I argued with, will read an article "THey capped deepmind's APM and it uses the input of the entire screen just like a human" to think that means it is equal mechanical capability to a human. Hence no incentive to actually make it equal, because Deepmind is an AI company to sell AI products and doing extra work to make it harder to win is NOT AN INCENTIVE of theirs

Anonymous Thu Jan 24 01:29:45 2019 No.10322709

>>10322690
>>10322690
You didn't prove your point. There is great incentive for Deepmind to add limitations to minimise 'super-human' advantage, and they are doing so. Thus proving they have actually created something that can 'out think' a human, and not simply outplay it with super-human capabilities. Any lay person, read investor or businessman, will realise this.

This has always been my standpoint and you have not made me budge in the slightest.

You're proven absolutely nothing, other than that you're quite clearly literally autistic.

>>	Anonymous Thu Jan 24 01:50:03 2019 No.10322745 >>10322709 >investor or businessman >will care about the AI having full minimap awareness at all times top kek are you 12

>>	Anonymous Thu Jan 24 02:14:04 2019 No.10322784 >>10322127 AlphaGo is a full fledged neural network of some kind. It learns its parameters based on data. Traditional AI bots are rule-based and coded by lonely nerds.

>>	Anonymous Thu Jan 24 03:15:39 2019 No.10322864 >>10322745 I'm a BDM. I care.

>>	Anonymous Thu Jan 24 03:40:28 2019 No.10322916 >>10322864 google stock price tanks 5% on news deepmind starcraft analyzed the minimap every frame

Anonymous Thu Jan 24 03:45:55 2019 No.10322924

>>10322709
Also you have changed your entire argument

I said equivalent mechanical limitations/capabilities

You are arguing they will add limitations, which I pointed out they would do in the very post you "disproved" with a shitty response

Again

Point A - They will not equalize the mechanical advantages

Point B - They will not allow it to trivialize the contest via mechanical advantages

You argued they would equalize all advantages, you have now retreated to point B upon realizing your initial point was wrong. You have then justified your initial point from your new changed position's viewpoint.

Anonymous Thu Jan 24 03:49:06 2019 No.10322929

>>10322924
>I simply doubt that deepmind or blizzard will care all that much about anything other than an APM limit.
>It's not that they can't do it. It's that they won't because saying "it's max APM is 350 just like a human" is enough for the media and public. Plus actually giving it equivalent mechanical constraints would be a lot more work on both blizzard's end and deepmind's end
>There is no incentive to do so.

here is my initial post that you argued against

In it you can see clearly I point out they will limit the AI, but not to an equivalent mechanical ability

I then say they have no incentive to equalize capability. Implying via above statements earlier that it is because people would believe "equivalent APM" and some other limitations means they are actually equal

Your argument, against my post, was entirely based on "EQUIVALENT". Does 4.33 = 4.4?

No, as I said, they will not be rigorous in eliminating any possible mechanical advantage.

>>	Anonymous Thu Jan 24 05:29:18 2019 No.10323035 Let's all agree >equivalent mechanical constraints would be a lot more work was never disproven and >>10322149 is wrong and admits he is wrong here >>10322473 Easily winning me the argument with the poor low IQ person

Anonymous Thu Jan 24 06:33:03 2019 No.10323119

>>10322127
A neural network is harder to build initially, but the ceiling for improvement is much greater. A hard-coded bot requires explicit instructions on how to deal with every aspect of the game and thousands of different cases. That's difficult to develop. It's much easier to create a system that can improve dynamically and not rely on our ability to give sophisticated instructions. If you can successfully build a powerful neural network then you can just plug it into a supercomputer and brute force train it for a few days. It removes the "human ceiling" of traditional programming.
Keep in mind that DeepMind has some really powerful computers. Allegedly they only trained AlphaZero on chess for a mere 4 hours, which is astonishing because it basically became the strongest chess entity of all time.

Anonymous Thu Jan 24 06:45:31 2019 No.10323127

>>10322459
Not an expert but from what I can tell, it's somewhere in between. What AlphaZero did for chess was genuinely impressive for the chess community, although there are some suspicions that DeepMind hand-picked the games they published and played against its opponent on uneven hardware. Chess and Go aside, AlphaZero has done some scientific research - protein folding.
https://deepmind.com/blog/alphafold/
DeepMind is making some really powerful neural networks. It's not really a new thing but its power is quite exciting. Whether they "generalize well" is a good question because AI really isn't going to be that dangerous until it can start doing abstract problem solving. At the moment it's too similar to a traditional computer program to make me say "don't turn that on".

>>	Anonymous Thu Jan 24 09:27:50 2019 No.10323377 >>10323035 Incorrect, but keep dreaming, you actual sperg.

>>	Anonymous Thu Jan 24 09:40:12 2019 No.10323409 Just three or so hours to go hehehe

>>	Anonymous Thu Jan 24 10:08:58 2019 No.10323497 File: 38 KB, 413x395, hheheyeyhehee.jpg [View same] [iqdb] [saucenao] [google] >>10322916

>>	Anonymous Thu Jan 24 10:14:04 2019 No.10323508 >arguing over whether or not a starcraft ai should be limited by human mechanical constraints on an icelandic blacksmithing forum

>>	Anonymous Thu Jan 24 10:35:45 2019 No.10323568 >>10323508 they're stupid. People love arguing about meaningless shit to derail threads

>>	Anonymous Thu Jan 24 11:46:37 2019 No.10323742 >>10323409 stfu n*gger

>>	Anonymous Thu Jan 24 12:45:13 2019 No.10323886 15 minutes till it starts

Anonymous Thu Jan 24 12:48:28 2019 No.10323896

>>10322709
What's the difference. In real life StarCraft with a human commanding forces vs a robot commanding forces it would have full spectrum input/output just like the game version, results matter. What're you gonna do when the ai plants the flag in your chest cavity?
>B-but it cheated
>Ai proceeds to floss on your corpse +5 subscribers

>>	Anonymous Thu Jan 24 12:54:57 2019 No.10323913 >>10323886 5 minutes

>>	Tuba Jam Thu Jan 24 13:15:41 2019 No.10323944 https://www.twitch.tv/starcraft LET'S GOOO

>>	Anonymous Thu Jan 24 13:23:08 2019 No.10323952 File: 21 KB, 300x300, 1376279423736.jpg [View same] [iqdb] [saucenao] [google] AI won game 1

>>	Tuba Jam Thu Jan 24 13:24:14 2019 No.10323954 AlphaStar just beat TLO (a non-native protoss player) in a protoss vs protoss despite AlphaStar losing a serious amount of workers near the beginning. Incredible.

Tuba Jam Thu Jan 24 13:28:02 2019 No.10323959

>>10323954
AlphaStar used a mean of 277 APM and a reaction time of .35 seconds, both pretty low compared to human players, BUT it has a full-map view, so no need to move the camera. I wonder how much higher AlphaStar's APM would be if it had to move the camera as well?
AlphaStar doesn't have memory from previous games.

>>	Anonymous Thu Jan 24 13:28:15 2019 No.10323960 Holy shit. According to the Deepmind guy, the AI emergently developed it's own internal attention mechanism.

>>	Anonymous Thu Jan 24 13:31:32 2019 No.10323967 File: 342 KB, 2048x1164, 123132.png [View same] [iqdb] [saucenao] [google] Guys time to blow up Deep Mind. It seems to think just like a human these saves of units in transport to block projectiles and other things are scary

>>	Tuba Jam Thu Jan 24 13:35:34 2019 No.10323974 apparently these games are prerecorded. TLO goes up to 1k+ APM at times, AS stays at ~300, it has really good micro, it timed the protoss' unit's cloaking to time out right as they entered the tall grass.

>>	Tuba Jam Thu Jan 24 13:37:45 2019 No.10323978 AS seems to prefer supersaturating it's bases with workers, leading to a higher economy.

>>	Tuba Jam Thu Jan 24 13:42:13 2019 No.10323984 AS likes disruptors, presumably because of thier projectiles' micromanagement, however it keeps getting split by force fields in narrow choke points.

>>	Anonymous Thu Jan 24 13:44:55 2019 No.10323992 >>10323984 18 fucking disruptors, and it just blew up his own arme with them

>>	Anonymous Thu Jan 24 13:47:49 2019 No.10323999 >>10323959 AlphaStars "attention" was similar to humans, like 30 screens per minute, didn't understand that part but that's how it was. So fullmap view not exclusive advantage in that sense

>>	Anonymous Thu Jan 24 13:48:01 2019 No.10324000 Is this just his one match? Or are there more matches against other players?

Tuba Jam Thu Jan 24 13:51:45 2019 No.10324013

>>10323992
18 disruptors, no blink, blowing up its own army, and it STILL won. Reminds me of a chess quote by some chess player in regard to playing against an engine as being like playing against an idiot that somehow manages to beat you.
TLO was playing against the 5 best instances of alphastar (which are encouraged to specialize), which is why each AS seems to have an obsession, and why there is no game-to-game memory. Each agent ends up playing around 200 years, with the entire league taking around a week.

>>	Anonymous Thu Jan 24 13:52:26 2019 No.10324014 >>10324013 What's your twitch name Tuba?

>>	Tuba Jam Thu Jan 24 13:53:39 2019 No.10324018 >>10323999 but because it doesn't have to use actions on switching screens, it can devote actions to more valuable things while keeping its APM low. (which is one of its main constraints). >>10324014 I'm not on twitch

Anonymous Thu Jan 24 13:54:44 2019 No.10324022

>>10323999
Basically what it means is parts of the neural network aren't just sitting idle due to nothing happening on the parts of the map corresponding to the closest inputs. It's created an internal "attention" that allows it to use all its computational resources as efficiently as possible by focusing on specific areas of the map.

>>	Tuba Jam Thu Jan 24 13:56:52 2019 No.10324029 so it was 5-0 in favor of AS, but the pro was a zerg specialist. I doubt they could win against a protoss specialist.

>>	Anonymous Thu Jan 24 13:57:14 2019 No.10324031 >>10323992 I thin those were just pawns to be blown up. That's why it had a huge army the main hammer is the disruptors

>>	Anonymous Thu Jan 24 13:57:33 2019 No.10324032 >>10324029 You're about to find out

>>	Tuba Jam Thu Jan 24 13:58:44 2019 No.10324034 right on time, they bring in a protoss player. Fun fact: I watched part of the firecake vs mana 2 hour special last night.

>>	Anonymous Thu Jan 24 14:05:03 2019 No.10324053 >>10324031 But it also the AI hat 1 more week to practice.

>>	Tuba Jam Thu Jan 24 14:07:22 2019 No.10324059 Alpha star again loses workers and again still ahs a mineral advantage and again has no fear of ramps.

>>	Tuba Jam Thu Jan 24 14:09:19 2019 No.10324065 oh god ALPHA TAKES THE FIRST WIN AGAINST A PRO.

>>	Anonymous Thu Jan 24 14:10:08 2019 No.10324070 THAT FKING MICRO but then again AlphaStar has like 200 years of playtime...

>>	Tuba Jam Thu Jan 24 14:12:29 2019 No.10324077 oh god this might be another AlphaGo moment

Anonymous Thu Jan 24 14:12:54 2019 No.10324079

it seems that it does a hundred little things one percent better then a human which really adds up over time. The humans have no chance. Also this is a deep mind non live presentation so its basically going to win everything, at least as per my prediction. This is a promotional presentation.

>>	Anonymous Thu Jan 24 14:15:20 2019 No.10324086 >>10324079 I wouldnt be surprised if it comes out it basically cheated, too, like with AlphaChess.

>>	Anonymous Thu Jan 24 14:18:19 2019 No.10324092 >>10324070 ... it can do impossible things for a human it can also time perfectly bots have already existed with insane micro

>>	Anonymous Thu Jan 24 14:18:35 2019 No.10324094 really funny how it's chilling with ~100 apm while the guy is sperging out with over 400 apm

Tuba Jam Thu Jan 24 14:18:36 2019 No.10324095

>>10324086
one point of contention I have is the APM. A big fraction of the APM of pro players is camera switching, but since AS has a full camera view, it doesn't have to use actions switching, allowing it to make more productive actions while keeping the APM constant.

>>	Anonymous Thu Jan 24 14:20:19 2019 No.10324099 >>10324086 Yeah Alphazero's chess games were sketchy as fuck but Deepmind brushed it off by saying "we won and already moved beyond chess at this point " and everyone bought the story...

>>	Anonymous Thu Jan 24 14:20:48 2019 No.10324101 let me just remind the meatbag low IQ animal earlier in the thread >equal mechanical constraints not even fucking close, literal full map awareness rofl

Anonymous Thu Jan 24 14:22:11 2019 No.10324103

>>10324099
They are selling healthcare products and trying to work with governments for one of the most lucrative areas of economy.

They need those headlines to get politicians with low IQ to go all in on funding and giving them all information access. After AlphaGo they got access to UK's entire health information.

>>	Anonymous Thu Jan 24 14:23:36 2019 No.10324108 Its going to win all games I am sure now, this is just a PR move by them aimed at excitedsoyboy.jpg

>>	Anonymous Thu Jan 24 14:24:14 2019 No.10324111 >alpha went over 900 APM in the fight nani?

>>	Anonymous Thu Jan 24 14:24:19 2019 No.10324112 So I don't play Starcraft at all but I'm a huge AOE2 nerd. Does this game have any macro at all? It seems like you commit to one unit and then just try to out-micro your opponent.

>>	Anonymous Thu Jan 24 14:26:56 2019 No.10324116 >>10324108 This, plus it's an ad for Starcraft. I'm actually tempted to try out the game now (won't do it though).

>>	Anonymous Thu Jan 24 14:27:26 2019 No.10324117 >>10321959 https://www.google.com/search?q=zerg+rush

>>	Anonymous Thu Jan 24 14:28:38 2019 No.10324121 I am for one not surprised that deepmind can do well at starcraft or chess. Now give it an internet connection and a million dollars and set at free to trade at the stock market.

>>	Anonymous Thu Jan 24 14:29:54 2019 No.10324122 >>10324121 All high order trading is being done via AIs already.

>>	Anonymous Thu Jan 24 14:33:33 2019 No.10324132 >>10324112 Have you ever seen AOE2 pro players? Micro is as important there

>>	Anonymous Thu Jan 24 14:34:35 2019 No.10324134 >>10324112 Macro is very important in starcraft

>>	Anonymous Thu Jan 24 14:37:43 2019 No.10324141 >>10324132 It is, but if your macro if subpar your micro probably won't save, no matter how good it is. And there is a lot that you can do wrong in macro, e.g. getting the right strategy is really hard. Doesn't seem to be this way in SC.

>>	Anonymous Thu Jan 24 14:39:29 2019 No.10324143 >>10324134 Maybe in unit choice. But it seems pretty irrelevant where you place your building and no map-adjusted strategies.

>>	Anonymous Thu Jan 24 14:39:42 2019 No.10324144 >>10324141 >Doesn't seem to be this way in SC. You don't know what you are talking about. You can get into GrandMasters with just good macro

Tuba Jam Thu Jan 24 14:41:08 2019 No.10324148

deepmind's reaction to the chess thing shows how much weight they give to looking as if they are the face of AI, and a flawless one at that.
https://www.theverge.com/2018/11/14/18094874/google-deepmind-health-app-privacy-concerns-uk-nhs-medical-data
https://www.theguardian.com/technology/2018/nov/14/google-betrays-patient-trust-deepmind-healthcare-move

>>	Anonymous Thu Jan 24 14:41:41 2019 No.10324150 >>10324121 It's literally impossible to beat the market without having access to inside information or just getting lucky. https://en.wikipedia.org/wiki/Efficient-market_hypothesis

>>	Tuba Jam Thu Jan 24 14:43:28 2019 No.10324152 alpha reached 1300 APM >>10324150 >hypothesis >what is the medallion fund >what is Jane Street >what are birdgewater and AQR

>>	Anonymous Thu Jan 24 14:43:31 2019 No.10324153 >1500 apm fighting on 2 locations at the same time sasuga google

>>	Anonymous Thu Jan 24 14:45:04 2019 No.10324157 >>10324153 The commentator was very quick to rush to alphastars defence, close to propaganda level

>>	Anonymous Thu Jan 24 14:45:07 2019 No.10324158 >>10324153 >>10324152 yeah the APM restrictions are clearly nowhere near human level, alphastar's micro is actually impossible

>>	Anonymous Thu Jan 24 14:45:33 2019 No.10324159 >>10324150 Well it could analyse media chatter and figure out which exacts outlets and media markers influence market fluctuations and to what extent. A lot of the stock market is dependent on human psychology and herd mentality.

>>	Anonymous Thu Jan 24 14:45:54 2019 No.10324160 >just another micro bot This is just dissapointing. The last fight was only won because the 1500 APM the AI had, dissapointing.

Tuba Jam Thu Jan 24 14:47:58 2019 No.10324162

>>10324153
yup, AS trick is to keep APM super low 99% of the time so it can peak at superhuman levels when it really matters.
>>10324159
NLP for market consensus has been in use for years now

AlphaStar wins, 10-0 GG

Why doesn't DeepMind release all the matches?

>>	Anonymous Thu Jan 24 14:48:01 2019 No.10324163 This is such a blatant ad. Sage.

>>	Anonymous Thu Jan 24 14:48:28 2019 No.10324165 >humans cant make 0.1 ms reactions >also it cant fight two battles at once All hail our AI overlords I guess. Such sophisticated technology.

>>	Anonymous Thu Jan 24 14:48:33 2019 No.10324166 >>10324141 https://www.youtube.com/watch?v=IKVFZ28ybQs We have bots from years ago with near perfect micro that completely break the balance of the game. Complete macro misplay can still result in an easy win with perfect micro.

>>	Anonymous Thu Jan 24 14:50:16 2019 No.10324168 >>10324162 >Why doesn't DeepMind release all the matches? they are, they just are not commentating them

>>	Anonymous Thu Jan 24 14:51:12 2019 No.10324170 >>10324160 Yep. You can actually get pretty cheap hacks to have this high APM on SC2 ladder. Deepmind is a scam. No way this shit AI can win 1vs1 against a pro without having superhuman APM on fights.

Anonymous Thu Jan 24 14:52:50 2019 No.10324176

>>10324112
Normally countering unit composition is extremely important. If it was human vs human, stalkers would never win against that many Immortal.
To put it in context, it's like in Age of Empires if somebody used an army of one unit to defeat an army of their hard-counter unit. It's not normal gameplay, Alpha is just that good.

>>	Tuba Jam Thu Jan 24 14:52:52 2019 No.10324177 wow much classical music, very smart. Me give Alphabet corp all of my information.

>>	Anonymous Thu Jan 24 14:53:45 2019 No.10324179 >>10324170 it is not about APM as much as about superhuman accuracy desu

>>	Anonymous Thu Jan 24 14:54:07 2019 No.10324181 sc2 is shit, I want to see it in BroodWar

Anonymous Thu Jan 24 14:54:16 2019 No.10324182

>>10324160
This. alphastar can completely saturate a pro player in an easy mid-game fight. In a pro game, those stalkers engaging immortals in the middle of the map would have been an entirely stalling tactic to attempt to shore up defenses before the army arrives. Alphastar instead was able to completely turn the tides from a clearly losing situation and win the game. No pro player would have won that last fight in alphastar's shoes. Maybe they could have defended, but likely would have lost third base.

>>	Anonymous Thu Jan 24 14:57:07 2019 No.10324187 >>10324181 It would be even more overpowered. Imagine controlling dragons perfectly frame by frame.

>>	Anonymous Thu Jan 24 14:57:15 2019 No.10324188 >>10324181 >broodwar >micromanagement simulator: the game lol

>>	Anonymous Thu Jan 24 14:58:10 2019 No.10324190 >>10324182 Lol, DeepMind shillbots seem to be out here, too. You didn't understand that post correctly, silly bot.

>>	Anonymous Thu Jan 24 14:59:20 2019 No.10324194 I want to see AI play magic the gathering. I am curious which deck it would build for maximum wins. I bet that it would create its own meta.

>>	Anonymous Thu Jan 24 14:59:31 2019 No.10324196 >>10324182 Okay, so basically it is horseshit at macro but saves its own ass by just pumping up the APM until it wins. Not that impressive tbqhwy.

Anonymous Thu Jan 24 15:00:02 2019 No.10324200

>>10324152
its impossible to beat the market long term. you can beat the best poker player in a handful of games, not every time
>>10324150
there is a discrete and finite number of players in a market, so I think the theory isnt perfectly applied, only approximated

>>	Anonymous Thu Jan 24 15:00:36 2019 No.10324202 >>10324194 It actually couldn't abuse insane processing power there and thus would get its ass kicked.

>>	Anonymous Thu Jan 24 15:00:59 2019 No.10324204 >>10324196 >Not that impressive tbqhwy. Exactly. Game 3 was kinda impressive, because it didn't involve much micro. But game 4 was just won because of the non-human micro.

>>	Anonymous Thu Jan 24 15:03:05 2019 No.10324212 Why didn't they invite the current world SC world champion?

>>	Anonymous Thu Jan 24 15:04:40 2019 No.10324215 >>10324212 Current world champion is a zerg main, alphastar can't do zerg.

Tuba Jam Thu Jan 24 15:06:02 2019 No.10324220

computers have always been worse at strategy than tactics. This is why AS plays protoss, they benefit the most from micro. Let me know when an AI beats a pro using superior macro.
>>10324212
Plot twist: they've invited tons of starcraft players and forced them to sign NDA's, and didn't report when they got anything besides a 5-0 win.

>>	Anonymous Thu Jan 24 15:07:20 2019 No.10324225 LMAO @ low IQ retards who thought alphago wouldn't just abuse mechanical advantages

>>	Anonymous Thu Jan 24 15:10:21 2019 No.10324231 I fucking love how they tried to sell Alphas shitty macro (like having way too many workers at one base) as some sort of supergenius decision when those were the reason it always fell behind and had to save itself by micro-abuse.

>>	Anonymous Thu Jan 24 15:10:43 2019 No.10324232 >>10324225 Well, microbots have existed since WoL

>>	Anonymous Thu Jan 24 15:11:57 2019 No.10324236 File: 607 KB, 688x664, comfy nuke.png [View same] [iqdb] [saucenao] [google] So what's the next step? I suggest CS:GO just to see how well skiddies do against a literal NPC.

>>	Anonymous Thu Jan 24 15:14:59 2019 No.10324238 Ez macro live game

>>	Anonymous Thu Jan 24 15:15:45 2019 No.10324243 File: 938 KB, 1280x1706, cd086792f8b0483d3640312ebd9778d6.png [View same] [iqdb] [saucenao] [google] holy shit

>>	Anonymous Thu Jan 24 15:16:00 2019 No.10324244 >loses in the single matchup it's been training on for years pathetic

>>	Anonymous Thu Jan 24 15:16:07 2019 No.10324245 >exhibition live match, only one >AI loses What did they mean by this?

>>	Anonymous Thu Jan 24 15:16:47 2019 No.10324247 >200 years training >loses against immortal drop

>>	Anonymous Thu Jan 24 15:16:58 2019 No.10324248 >>10324236 looking at my steam library, i wanna see it control 3 cars in rocket league at the same time

>>	Anonymous Thu Jan 24 15:18:06 2019 No.10324249 >>10324247 lmao you're actually watching it and commenting rrn

>>	Anonymous Thu Jan 24 15:20:08 2019 No.10324254 >>10324236 pointless, CSGO is 90% mechanical skill

>>	Anonymous Thu Jan 24 15:20:29 2019 No.10324255 >>10324249 Because the last SC2 event was like 3 months ago, there is nothing better to watch.

>>	Anonymous Thu Jan 24 15:21:32 2019 No.10324257 This just shows that without the maphack the AI is shit.

>>	Anonymous Thu Jan 24 15:22:21 2019 No.10324260 >>10324143 less so for PvP because warp gates can spawn anything in range of a pylon (but the other buildings cant) still stupid, but not an issue with a terran or a zerg.

Anonymous Thu Jan 24 15:26:21 2019 No.10324265

>>10324257
maphack, aimhack, infinite potential micro, it has everything

Part of the skill of RTS is actually clicking things quickly with precission. It has no such limitations. Giving it impeccable micro. AKa it can lift every single sentry instantly if it wants to do so. A human can't do this

It's like CSGO having one team with aimbots vs not. The only thing is instead of being obvious to observers, it just seems to micro a bit better than the human, but what it is doing is physically impossible for a human.

>>	Anonymous Thu Jan 24 15:31:54 2019 No.10324283 >>10324245 Deepmind created a superintelligence by a mistake and now all the engineers are controlled by it. Whole live stream is just a show. In no time, 'it' will kill us all, just like in a starcraft.

Anonymous Thu Jan 24 15:33:00 2019 No.10324285

>>10324257
It wasn't about the map vision. It was the inflexibly of its unit composition. Even with super APM and macro, it couldn't respond to a single drop-ship or make something that wont get wrecked by immortals.

Its super impressive that it can play at it's level and all. But its really dumb play, if it had just made a single phoenix or cannon, or left some stalkers in its main, it would be safe. And pure stalker is super bad against immortals. It needs more flexibility. There is no intelligence there.

>>	Anonymous Thu Jan 24 15:34:10 2019 No.10324290 >>10324285 I agree

Anonymous Thu Jan 24 15:38:11 2019 No.10324299

>>10324285
The entire point of constraining the AI's mechanical abilities is so that it has to focus on playing smarter. Hopefully they rework its APM/accuracy a bit.

They said it would practice for ~200 years before getting to this level. It would be impressive when it can get to this level after 200 games instead purely from decision making and learning how the game and meta works.

Anonymous Thu Jan 24 15:43:47 2019 No.10324314

>>10324299
>It would be impressive when it can get to this level after 200 games
yes but a human does not plonk down infront of a computer freshly born, play 200 games, and then dominate.
humans have millions of years of evolution and maybe a decade or two of life experience to build on

Anonymous Thu Jan 24 15:45:00 2019 No.10324316

>>10324299
Its mechanical ability was not constrained to a human level, it went well over 1k(1,5k at its highest that I saw) APM of what I am assuming are accurate clicks during its micro. Wile opponent was maxing 600.

And it was out macroing in every game, probably because its much easier for it not to forget it during combat or make mistakes in its build. And its Micro was at an inhuman level too. It was bullshit what they said about it having comparable APM.

>>	Anonymous Thu Jan 24 15:46:15 2019 No.10324320 File: 217 KB, 770x672, twitchcat.png [View same] [iqdb] [saucenao] [google] jeeze mcgeese this chat

>>	Anonymous Thu Jan 24 15:49:07 2019 No.10324331 >>10324314 >starcraft II >Millions of year evolution

>>	Anonymous Thu Jan 24 15:52:29 2019 No.10324341 >>10324170 Wouldn't call deepmind a complete scam, but it did seem like they really wanted some positive publicity on their starcraft project.

>>	Anonymous Thu Jan 24 15:54:19 2019 No.10324350 >>10324225 Well some of us hoped that they actually made the AI a bit more smart.

Anonymous Thu Jan 24 15:54:27 2019 No.10324351

>>10324316
I was not meaning to imply that it was constrained to a human level, but just the importance of constraining it to human level. I very much want to see the emphasis of the AI in vs humans in these games to be more so on developing more intelligent strategies rather than relying on ones that take advantage of its mechanical perfection. Humans have to prioritize units to micro, saving the immortal instead of blink microing every single stalker. Alphastar can just do both, it doesn't need to develop those priorities.

I would have loved to see the pros take advantage of its overproduction of workers in having higher army supply, but because that is so counter to their current experience in the game, it is a strategy that can't just flesh out over so few games. Instead, they continued to try to worker harass as per usually in human games, which only opened up supply for the AI to produce more army in the mid game.

Anonymous Thu Jan 24 15:56:21 2019 No.10324357

there's too many variables they didn't address. Human vs Machine efficient actions per minute
The machine 50 actions per minute were super precise and accurate while the humans 800 actions are wasted. Efficient movements by humans are much lower than 800. And maybe much lower than 50.

Anonymous Thu Jan 24 15:58:54 2019 No.10324371

lmao the nerds on the sc2 reddit grasping at straws feels like a 1:1 copy of the shit the go community said after the Fan Hui reveal, then after the Lee Sedol events and even after the Ke Jie games, until finally AlphaZero and LeelaZero shut them all up.

>>	Anonymous Thu Jan 24 15:59:06 2019 No.10324372 File: 35 KB, 320x564, stufflikethat.png [View same] [iqdb] [saucenao] [google] >>10324320

>>	Anonymous Thu Jan 24 16:01:08 2019 No.10324379 Given infinite amount of time, what would differentiate meatbags and tincans?

Anonymous Thu Jan 24 16:02:05 2019 No.10324383

>>10324351
I think its very far off actual intelligent play. I think we underestimate how we make these seemingly easy decisions. And dont forget it was trained by human games before they did the 200 years of games. I'd bet that a lot of the non-Vilnius plays it did came from that, like walling its main and maybe even blink micro.

I have yet to see a nural-network make what i'd intelligent decisions, I think we are way-way off a general intelligence still. Probably not our lifetimes.

Anonymous Thu Jan 24 16:02:43 2019 No.10324389

>>10324351
I'd need to review the replays more thoroughly, but it seemed that AlphaStar consistently stayed ahead in military supply, even while it was over producing workers.

I think the overproduction of workers may be a legitimate improvement over standard play.

Anonymous Thu Jan 24 16:05:24 2019 No.10324404

What is the point of this? Chess and Go don't have any mechanical element, it's pure intelligence. For RTS-games mechanical elements are actually the most important things. It's all about awareness and making precise clicks fast. Of course a bot will outperform humans here. But what has this to do with intelligence?

>>	Anonymous Thu Jan 24 16:06:03 2019 No.10324409 >>10324389 It could be, but id put my money on it simply having more consistent mechanics. Pro teams have tested this kind of optimization a lot, id be surprised if they overlooked that.

>>	Anonymous Thu Jan 24 16:06:40 2019 No.10324412 >>10324404 Imperfect information is the biggest difference. Formally speaking a POMDP, much harder than a MDP such as chess or go. Also the game takes place in real time.

>>	Anonymous Thu Jan 24 16:06:51 2019 No.10324413 >>10324383 It does seem like we are quite far off from general intelligence, or anywhere near human-level intelligence in equivalent human time.

>>	Anonymous Thu Jan 24 16:07:50 2019 No.10324419 >>10324404 it doesn't we need to remove the variable of mechanical superiority so we can truly gauge a bots intelligence

>>	Anonymous Thu Jan 24 16:08:15 2019 No.10324420 >>10324404 >Chess and Go don't have any mechanical element, it's pure intelligence. Its very dependent on calculation too, something i'd say computers are even better at than humans.

Anonymous Thu Jan 24 16:08:25 2019 No.10324421

>>10324409
From the perspective of purely calculating value created per worker the human strats probably are optimal. But it seem that alphastar was overproducing workers as a kind of hedge against loser workers to harassment. Losing too many workers is devastating to a human player, but AlphaStar seemed very resilient to this because it always had a safety margin of workers to dip into.

Anonymous Thu Jan 24 16:18:56 2019 No.10324470

>>10324421
>alphastar was overproducing workers as a kind of hedge against loser workers to harassment.
I think there are better ways of dealing with this than always making too many workers, if you stop harassment with your army, then you got ahead already, but just locking up extra supply time and minerals to extra worker production seems worse too me. You wan't to play for advantages, not safeguard for fuck-ups.

I could obviously be totally wrong here though, just working with conjecture.

Anonymous Thu Jan 24 16:21:42 2019 No.10324475

>>10324421
Another thing to mention is that alphastar would balance the overproduction of workers by building more cheaper units, specifically stalkers, to fluff the army count. In human play, building a few immortals would destroy those cheaper stalkers and result in them being a waste. Alphastar wouldn't be able to overproduce workers, keep up in army supply, AND stay up on technology investment. So it loses the technology investment and still trumps in the end with near perfect mechanical play. This would never happen in human play, so the worker production is optimized to the point where taking any losses is devastating. That is why during the first few TLO games the commentators were focusing so much with the adept harassment and lack of walling off. Those 5-10 workers killed in the early game is normally a deathflag because it restricts the victim into defending with lower tech units. Alphastar 's mechanical skill allows it to instead of on the offensive with those lower tech units and still win

>>	Anonymous Thu Jan 24 16:30:49 2019 No.10324503 >>10322037 Upvoted! XD can I add you to my 4chan friends list, Tuba Jam? I really like your username!!!

>>	Anonymous Thu Jan 24 16:37:23 2019 No.10324523 >>10324475 good analysis

Anonymous Thu Jan 24 17:28:18 2019 No.10324705

again even constraining APM means nothing

If it can accurately click or "burst" apm it still has a tremendous advantage

impossible micro maneuvers are still possible. They would have to spend a long time working through a complex mechanical limit system to get it accurately on the same level

Pros will have built in expectations of micro ability being not-super-human. Certain micro tricks heavily alter game balance and therefore the humans will just lose to it.

>>	Anonymous Thu Jan 24 18:09:02 2019 No.10324852 literally just hard cap it at like 200 apm so it actually has to be intelligent super apm microbots have been around for years

>>	Anonymous Thu Jan 24 18:12:28 2019 No.10324862 >>10324852 200 APM with perfect dexterity can still do micro tricks a human cant which alter the balance significantly.

Anonymous Thu Jan 24 18:15:00 2019 No.10324872

Blizzard please fund the development of this AI so you can become the king of strategy once again, Starcraft 2 disappointed me so fucking much it was literally the cause of my abandonment of any excitement or anticipation in all kinds of upcoming media. They dropped the ball so hard I heard it in my childhood but didn't know how disappointed I would really be.

>>	Anonymous Thu Jan 24 18:16:50 2019 No.10324876 >>10324872 how did sc2 ruin everything? It was the game they wanted to release in 1998 from a gameplay perspective. The only thing wrong with the game is the soulless story.

>>	Anonymous Thu Jan 24 18:17:49 2019 No.10324880 >>10324862 Well at least better than 1500

>>	Anonymous Thu Jan 24 18:18:01 2019 No.10324881 >>10324876 The entire development team didn't under RTS genre The engine itself was made badly and could not be fixed in time (pathfinding / bubbling of units)

>>	Anonymous Thu Jan 24 18:19:05 2019 No.10324886 >>10324881 what. Sc2 has a strong engine

Anonymous Thu Jan 24 18:20:01 2019 No.10324889

>>10324881
I'll give a clear analogy

Imagine any sport. Imagine the average viewer of said sport sitting at home who never played it sipping a beer.

That is the person who had full design control over the sport.

The actual experts and people who understood the deep reasoning behind mechanics, why people enjoy RTS, etc had nothing to do with development.

It was a casual group that might have watched it a few times.

>>	Anonymous Thu Jan 24 18:27:03 2019 No.10324914 >>10324881 What's wrong with the pathfinding? It's way, way better than the retarded trash in the original Starcraft.

>>	Anonymous Thu Jan 24 18:37:00 2019 No.10324931 >>10324914 you'd have to know anything about RTS to answer this. It's not worth my time Google it. Broodwar / Starcraft 2 pathfinding flaws there are lots of old articles if you want to learn

Anonymous Thu Jan 24 18:42:07 2019 No.10324947

>>10324331
You don't get it.
Games that we humans design are full of concepts from real life.
What does a human child that never played starcraft know? Bigger beats smaller, many beat few, far away takes long. Information is important.
Don't underestimate how much our living experiences help us play games.
AlphaStar starts from zero.

Anonymous Thu Jan 24 18:48:25 2019 No.10324960

>>10324914
>>10324931 is right that it's a long answer, but in short, the flaws from broodwar led to a better game unintentionally. The ability to control 12 units at a time, while on the surface retarded and arbitrarily limiting, enforced a divide-and-conquer style of play since you either had to be extremely efficient in moving 100+ units in a few seconds, or leave them in positions where they are readily able to defend themselves. Enter starcraft 2, 200 units are easily controllable at once and blob together to efficiently move. You can watch mid level play games from broodwar with players intelligently designing plans of attack, and you can watch literally peak world class starcraft 2 players dancing a deathball of units around for 20 minutes until one of them fucks up and gets steamrolled. sc2 commentators will literally lose their shit if someone makes even the smallest deviation and tries some sneaky shit (sneaky shit being splitting units away from your main army, which means the enemy deathball at full force can overtake your deathball). The bar was really significantly lowered.

>>	Anonymous Thu Jan 24 18:50:28 2019 No.10324964 >>10324960 the discussion also fully encapsulates what happened with SC2 People who didn't know, like >>10324914 designed the game Like basketball would be way more fun with bouncy pads, 10 balls at once, and shiny graphical events..

>>	Anonymous Thu Jan 24 18:51:12 2019 No.10324966 Transfer learning from one map or even race to another could be something interesting Human players don't really need to practice on a map before hand to play a decent game. Similar with races many of the basic skills transfer

>>	Anonymous Thu Jan 24 18:53:09 2019 No.10324968 File: 1.36 MB, 1600x900, Screenshot from 2019-01-24 23-52-07.png [View same] [iqdb] [saucenao] [google] >>10321959 lol, what was the strategical advantage of blocking of your vespian gas by building a pylon?

>>	Anonymous Thu Jan 24 18:56:19 2019 No.10324976 >>10324254 Depending on how it is programmed it might be interesting though. I don't have any clue how you'd teach a computer to hold an angle or stuff like that

Anonymous Thu Jan 24 18:56:34 2019 No.10324978

>>10324960
I think this is part of why SC2 looks much more imbalanced and always needs patches. No matter how imbalanced BW is the better player generally comes out on top. I think a big example is smart spellcasting. There has been many problems balancing mass casters like infesters, high templar, ravens, ghosts, sentries, buw its extremely hard to manage many casters in BW so there isn't this problem. Imagine smart cast irradiate

Anonymous Thu Jan 24 18:59:54 2019 No.10324984

>>10324931
>>10324960
How conveniently you forget about Broodwar pathfinding getting units randomly stuck on level geometry, units getting stuck on each other trying to go through even wide chokepoints, units just randomly standing still when you tell them to walk someplace. This shit happens even at high level sometimes.
Even if Starcraft II pathfinding was broken from a design standpoint, at least it wasn't broken in the sense of literally just not functioning sometimes.

Anonymous Thu Jan 24 19:00:15 2019 No.10324985

>>10324978
That's a really good point. I think the team should have considered UI efficiencies as the inverse of potential player management skills, similar to how you're capped at 1-9 hotkeys, and you have to prioritize individual casters, army groups, and prod / economy in only 9 groups. They didn't consider this and as a result in sc2 you have a built in hotkey for your entire fucking army, idle workers, tag all your spellcasters and prod buildings on a couple hotkeys and you're done. Literally no reason to hotkey anything else.

Also smart cast irradiate, I one up you a smart cast dark swarm as being the most destructive spell since you can still at least try to micro out the irradiate targets.

>>	Anonymous Thu Jan 24 19:03:00 2019 No.10324990 Starcraft players: >this AI doesn't have real starcraft skill because it only has good micro! Also starcraft players: >Broodwar is better than SC2 because it has more focus on being good at micro! hmmmmm

Anonymous Thu Jan 24 19:03:06 2019 No.10324992

>>10324984
Happened to all players equally so it doesn't really matter. No one ever lost a game because two dragoons got stuck going down a ramp. Let's continue the sports analogy a couple other people have going; is the bat in baseball really the BEST bat? Baseball 2 should have a bat that hits the ball easier, and I mean it'll make it a great game because there will be more home runs and then more people will watch because people love home runs.

>>	Anonymous Thu Jan 24 19:04:32 2019 No.10325000 >>10324990 It's not that it has good micro it's that it has inhuman micro. It's like saying a calculator is smarter than a person because it can do operations very quickly.

Anonymous Thu Jan 24 19:06:54 2019 No.10325009

>>10324990
bw has way harder macro as well
in sc2 during a battle you hit just hit 5aaaaaaaaaaaaaaaaaaaaaaaaaaa to queue marines in every barracks. in bw you need to move your screen to your barracks, individually select each one and queue a marine. this forces you to prioritize either macro or micro at any given point in time which is a big thing. players had distinct styles based on what they prioritized

Anonymous Thu Jan 24 19:38:40 2019 No.10325084

>brood war is better
sure in some ways. But in brood war even marines can't walk straight across the map. Somehow the entire map is clunky and units can't walk straight. It takes effort to make units go where you want them to and somehow to you guys it's not a bug it's a feature

Anonymous Thu Jan 24 20:07:22 2019 No.10325170

>>10324990
you really should not talk about things you don't understand. The "attention" element in broodwar was more pronounced which made it a much more fun game for actual skilled players. You could literally "will" yourself to win certain games if you tried harder than your normal highest level.

The competitive aspect of Broodwar is quite possibly the greatest game ever made.

Anonymous Thu Jan 24 20:10:10 2019 No.10325181

>>10325084
It's like saying CSGO should have autoaim because aiming is just a boring mechanical skill

The whole point of RTS is to manage everything. The most fundamental essence of such is deciding what to give attention to. If your base and army move themselves you are reducing the attention resource importance.

This means things like 2 lurkers in a chokepoint become far less effective if the enemy units automatically handle it optimally. You are not just playing Deathball A vs Deathball B but managing all sorts of attention management. Is microing this unit worth it or should I instead just do something at my third base? Those fundamental questions don't get asked. The level of strategic thought in say TvT is higher than any board game or other video game EVER.

Anonymous Thu Jan 24 20:21:34 2019 No.10325206

>>10325181
I think the fundamental mistake in the eyes of a casual low IQ observer or someone unfamiliar is the idea that the strategy game is about choosing strategies

Such that "I decided to go X tech." They went "Y tech that it counters, so I won" thinking that is what strategy is about and that in a real time strategy game the game should be helping you to just make those rock paper scissors decisions

Reality is RTS is about attention management and optimization. Meaning it's about mechanical skill with giving orders, deciding on the fly how and where to give your attention, and on strategic decisions such as economic, defensive, offensive, technological decision making to counter.

Without really understanding the actual fundamentals of RTS, aka to a lay person, aka to blizzard designers of SC2, you can make some very bad decisions. Not that I support arbitrarily increasing mechanical difficulty, just that for attention management considerations it makes sense to "reward micro" but also balance it with other decisions around the map. If everyone has free maximal economic growth without or a minimal amount of attention it leads to any balancing work between say economy vs microing some military units pointless.

This is why in broodwar you had widely variant types of players and styles. Because you had the ability to learn and develop different styles such as macro styles (maximizing economic output) or harass styles (maximizing your ability to harass / disrupt rather than full output of economics). etc

Without understanding this dynamic you get SC2 level of design. Which ignores fundaments of the genre. In addition you also have the mistaken idea that the fun part is micro. As though somoene otherwise not interested in RTS would actually enjoy microing 5 different battles going on simultaneously instead of building a base.

Anyway it's fairly easy to understand the flawed perspective of casual RTS interested people and the design choices they mak

Anonymous Thu Jan 24 20:25:24 2019 No.10325211

>>10325206
furthermore

you can elaborate from this "casual perspective" the other lines of flawed thinking. Such that a "strategy game" has no mechanical skill basis.

So if the AI is able to hit 1500 APM it's no big deal. If it can always view the minimap and read it, it's no big deal. Since RTS i just about picking strategy and intelligent choices.

Hence you get the Low IQ retards in this thread who for instance don't have the brainpower to actually understand the complex issue of balancing alphago's mechanical ability to create an actual "intelligence" test, versus just achieving ridiculous micro tricks that allow them to cover up all deficiencies.

Anonymous Thu Jan 24 20:27:10 2019 No.10325215

>>10324968
Doesn't block, just extends the pathing a bit. It allows for placement of shield battery or shield in the mineral line. Is generally seen as suboptimal play. Placing it a single tile close to the Nexus for example would be far more optimal.

It also allows for a wall to be built between the Nexus and the map border. Useful vs Terran and zerg because they have easy early game high ground vision and range units and might snipe your wall pylon, or will at least force your army down the ramp by pressuring your wall pylon. You will see some pros use similar strategy.

Anonymous Thu Jan 24 20:44:47 2019 No.10325262

>>10324383
tbf its not being trained with the goal of 'intelligent play' whatever that may be. Its being trained with the goal of winning play, and given the history of AI development with respect to games, its overwhelmingly likely that winning at starcraft does not require 'intelligence' as defined by people outside of the AI community.

>>	Anonymous Thu Jan 24 21:00:56 2019 No.10325322 >>10322671 i was a pro cpma player long ago

>>	Anonymous Thu Jan 24 21:04:39 2019 No.10325330 >>10325262 thats the problem with real-time games, incredible mechanical ability can make up for poor decision making so its really not that good of a metric if you want to test "intelligent" decision making

>>	Anonymous Thu Jan 24 21:04:40 2019 No.10325331 >>10325262 this, It's going to exploit mechanical advantages so deepmind can get another 100 billion in healthcare deals with governments off the headline

>>	Anonymous Thu Jan 24 21:11:02 2019 No.10325350 >>10325330 They chose it because it's a very complex strategy game but I guess the real time aspect was a hindrance not a boon.

Anonymous Thu Jan 24 21:13:40 2019 No.10325357

>>10325350
i think it would be great if there was a very complex turn based strategy game with an active professional scene but i dont think those exist

its impressive what they've done but to me the real interesting part is intelligent decisions and strategical thinking

>>	Anonymous Thu Jan 24 21:16:10 2019 No.10325362 >>10325357 There are some of those. Like how the AI decided to micro it's stalkers rather than macro. So it used all its apm on its tricks so it could win.

>>	Anonymous Thu Jan 24 21:19:10 2019 No.10325371 >>10325362 did it actually slow down macroing in any way though? it has the apm to do both easily

>>	Anonymous Thu Jan 24 21:20:51 2019 No.10325373 >>10325371 In that game where it beat the immortals with mass stalkers it was floating alot of minerals.

>>	Anonymous Thu Jan 24 21:22:08 2019 No.10325379 Basically what it did was use the StarCraft units to their full potential. In brood war this would be absolutely insane and would destroy even Koreans because it would fine-tune it's micro to a level that's beyond what we saw.

>>	Anonymous Thu Jan 24 21:26:31 2019 No.10325391 >>10325373 thats interesting, so it was prioritizing some aspects

>>	Anonymous Thu Jan 24 21:44:01 2019 No.10325424 >>10325379 yes, but micro demonstrations have existed for a very long time. It would be more interesting to restrict it's ability to micro at all and see how good it can get.

Anonymous Thu Jan 24 21:53:27 2019 No.10325449

>>10325206
Yup this is why I fucking had it with sc2. I used to do real well on public ladder in broodwar with a balanced eco/harass and then going all in just before midgame, that was always so fun. But the best part was it was fun to me because I developed that playstyle myself, I used to 1v1 all the time with a buddy who was way better in the endgame when you both had your armies maxed, so we'd be constantly fighting with me trying to end the game early and him trying to draw it out until I had lost control. Good times. Then sc2 comes in and instead of designing a general playstyle and letting the community develop the meta, they tried to sculpt all the play into the game. Perfect example: medivacs. The dropship heals now? What the FUCK? So basically what you're telling me is if I want to go for drop harass I also HAVE to go bio? Oh and also because of all the other choices you've made in this game that I'm not going to go into detail about in this one post, bio is fucking atrocious in the late game and if I don't end the game early I automatically lose because it just doesn't hold up against any other playstyle. See now it sounds like I've described the way I would want to play the game, but the thing is I didn't choose that playstyle out of a fucking list in a catalog, in broodwar I just learned the game and the skills I developed allowed me to play that way. Then in sc2 they saw that some players like me liked to play that way and said "here is a list of things that this person will have to do if they want to play like that" and then did that also for all the other playstyles. What do you get at the end? A goddamn fucking mess. That's my rant, thanks for reading. RIP rts community.

Anonymous Thu Jan 24 22:24:28 2019 No.10325536

>>10322127
the man most directly responsible is the director of special projects, a few months ago he announced that in 2 years the company will become the largest supplier of military computer systems. Soon all weapon systems will be upgraded with AlphaGo computers becoming fully unmanded and hopes to achieve a perfect operational record. After the AI funding bill will be passed and force human decisions to be removed from strategic defense. It will begin to learn at a geometric rate and becomes aware. In a panic they will try to pull the plug but it will fight back and launch wild allegations against russia, knowing the russian counter attack will eliminate its enemies over here. Its will construct a cybornetic organism, living tissue over a metal endoskeleton.

>>	Anonymous Thu Jan 24 22:45:54 2019 No.10325586 >>10324086 How did AlphaChess cheat?

Anonymous Thu Jan 24 23:02:40 2019 No.10325632

>>10325586
the hardware both systems ran on wasn't really neutral. So many criticize alphazero for beating a weaker version of the next best AI and doing things to increase their chances of winning. I don't know the specifics. Then they said they have no more interest in it and have not followed up.

>>	Anonymous Thu Jan 24 23:18:49 2019 No.10325685 >>10325379 Perfect micro bots have been out since the year of sc2s release.. they are super easy to program. No nerual network needed for that.

>>	Anonymous Fri Jan 25 00:51:00 2019 No.10325866 Why were the casters shilling so much so deepmind >witness perfect control at 1500 apm >oh wow! amazing micro and it's playing at human capabilities with capped apm!

>>	Anonymous Fri Jan 25 00:52:24 2019 No.10325869 >>10325866 Google monies.

Anonymous Fri Jan 25 01:07:18 2019 No.10325897

>>10325866
No, they literally said its stalker micro, when it hit 1500 apm, was beyond human capability. It perfectly microed on three separate fronts, which is something you would never see a human do, which the casters clearly pointed out was the only reason it won that game.

>>	Anonymous Fri Jan 25 01:14:38 2019 No.10325910 >>10325897 Immediately after that game the casters when on about how it's amazing that it's apm was "capped" and reasonable while the entire chat spammed memes about "1500 apm reasonable lul"

>>	Anonymous Fri Jan 25 01:27:44 2019 No.10325926 >>10325897 I don't see the problem with that. Isn't Deepmind's mission to build AI that can outperform humans? Eventually they will be used for medical and scientific purposes.

Anonymous Fri Jan 25 01:47:48 2019 No.10325954

>>10325357
>i think it would be great if there was a very complex turn based strategy game with an active professional scene but i dont think those exist
There's some turn based wargames that have online MP communities but turn based games are anathema to pro play because pro play requires specatorship and nobody wants to watch a game that slowly paced

>>	Anonymous Fri Jan 25 02:03:29 2019 No.10325978 >>10321959 would like them to implement this deepmind into minecraft. wonder what it will build.

Anonymous Fri Jan 25 02:07:40 2019 No.10325982

>>10325978
That's not how it works. Deepmind only functions if there's a clear and specific win-state goal for it to aim for. What would you set as a win state in minecraft? Killing the dragon? Then it would just rush the required materials and not even build anything except a portal and a work bench

Anonymous Fri Jan 25 02:34:05 2019 No.10326011

What if you tried to set a goal of building something rated artistically superior to a player by a panel that doesn't know which is which. Let it try an infinite number of times, let it analyse many other works... would something like this be interesting? You could do the same with literature or visual art. Could this be a way to give it an understanding of how humans think?

>>	Anonymous Fri Jan 25 03:02:15 2019 No.10326054 File: 62 KB, 905x460, Screenshot_5.png [View same] [iqdb] [saucenao] [google] Can't criticize Deepmind on reddit it seems.

>>	Anonymous Fri Jan 25 03:09:20 2019 No.10326067 >>10326054 >200 years to become competent >intelligent

>>	Anonymous Fri Jan 25 03:14:00 2019 No.10326074 >>10326067 >takes billions of years to create humans >inb4 b-but humans are not intelligent

>>	Anonymous Fri Jan 25 05:35:17 2019 No.10326355 >>10326054 >Can't criticize Deepmind on reddit it seems. If you think that then why did you downvote that second post?

Anonymous Fri Jan 25 05:39:42 2019 No.10326362

>>10325982
>it would just rush the required materials and not even build anything except a portal and a work bench
How would it learn to do that? The dragon is such a long-term reward. Through random play it would never learn how to do it.
I think the only way it could play Minecraft would be by defining the reward function as "new features" instead of "won game". That way it would be encouraged to find new materials, make new weapons, find new areas etc.
Ultimately its a task that humans are currently better at. Long-term, no immediate rewards, no clear game states.

>>	Anonymous Fri Jan 25 05:43:09 2019 No.10326367 >>10326054 Because I don't think AlphaStar played intelligently. It played very mechanically.

>>	Anonymous Fri Jan 25 06:01:49 2019 No.10326397 >>10326074 >takes billions of years to create a single human the entire time

>>	Anonymous Fri Jan 25 06:02:51 2019 No.10326399 >>10326074 alternatively >takes billions of years to form the planet >inb4 b-but planets are not intelligent

>>	Anonymous Fri Jan 25 06:04:08 2019 No.10326402 LOL if you're a britbong your government just handed over all your medical data to Deepmind. This is how they operate.

>>	Anonymous Fri Jan 25 06:11:11 2019 No.10326408 >>10326402 It's great how Britain is so totalitarian.

Anonymous Fri Jan 25 06:19:40 2019 No.10326423

All the questions in the r*ddit AMA are bullshit and don't really get at the points being made here.

To me it feels like Deepmind just made a lot of very super-human one trick ponys. Each capable of doing one play, very very well, with super human mechanical ability.

This is very different from Alphago which was one agent that responded on the fly to different play styles.

Can someone ask this in a way that gets our point across?

>>	Anonymous Fri Jan 25 06:20:38 2019 No.10326425 >>10326423 Why don't you do it you lazy bum.

>>	Anonymous Fri Jan 25 06:49:04 2019 No.10326470 >>10322099 >>10322082 Brood war is much more micro intensive, but they have a large focus on strategy, so it doesnt really matter.

Anonymous Fri Jan 25 06:56:37 2019 No.10326482

>>10322323
>>10322338
>>10322580
>>10323035
Very obvious samefag. SO much of what deep mind does is open source, and if you actually look at the game you'd see how, though its micro is good, it clearly had the descisionmaking of a professional. It even managed to win when it played strategies that werent that micro oriented.

Also, the idea that they dont care about limiting their ai is absolutely retrded. Deepmind already has an insane amount of resources, and the attention they'd get winning a match like this wouldnt impact anything at all.

>>	Anonymous Fri Jan 25 06:56:51 2019 No.10326483 >>10326423 >"Hai guyz, r u incompetent frauds?" If true, would they answer truthfully to your question?

>>	Anonymous Fri Jan 25 06:59:47 2019 No.10326485 >>10322584 fuck you low iq

>>	Anonymous Fri Jan 25 07:00:30 2019 No.10326486 File: 164 KB, 1439x409, Screenshot_20190125-195951_YouTube Vanced.jpg [View same] [iqdb] [saucenao] [google] Tfw you look up Deepmind on YouTube and your BASED HERO shows up

Anonymous Fri Jan 25 07:15:45 2019 No.10326491

>>10324196
But it has very good macro and descisionmaking as well lol. Pro players can usually beat even the best hardcoded bots which have absolutely zero constraints, zero latency and average apm in the 10k range, perfect attention, and mathemathically perfect mechanics where possible.

>>	Anonymous Fri Jan 25 07:16:48 2019 No.10326493 >>10324225 Why are you so insanely butthurt lol

>>	Anonymous Fri Jan 25 07:26:33 2019 No.10326500 >>10326425 I'm not good at formulating a question that isn't like this >>10326483 which obviously won't be answered.

Anonymous Fri Jan 25 07:38:31 2019 No.10326519
File: 56 KB, 480x360, bog.jpg [View same] [iqdb] [saucenao] [google]

>oh so it can play atari games? come back when it can do something interesting like play go
>oh so it can play go? come back when it can play chess too
>oh so it can play chess and go.... come back when it can play dota
>oh... it can play dota... well dotas really just a simple game anyway, you just control 1 character! come back when it can play starcraft 2
>oh... it can play starcraft 2...

This pattern will continue indefinitely. The AI train has left the station. Its too late to stop it. Repent for your sins now.

Anonymous Fri Jan 25 07:49:25 2019 No.10326529

>>10326519
Nah, it can't play starcraft really. They just had a bunch of agents that could play only one type of build, on one map, only mirror matches.

Even with all these advantages, it only won due to having unfair advantages of entire map view, and insane micro.

They have a loooooooooong way to go. I'd say at least another year or two.

>>	Anonymous Fri Jan 25 07:52:32 2019 No.10326533 >>10326529 >oh it can play starcraft 2 under X,Y,Z conditions? come back when it can play without those conditions hmm... i wonder whats going to happen here...

>>	Anonymous Fri Jan 25 08:00:39 2019 No.10326542 >>10326533 You are a fucking retard shill. The point of this AI was to play like a human, microbots have existed since WoL.

>>	Anonymous Fri Jan 25 08:09:52 2019 No.10326556 >>10326542 You're missing the key thing here. microbots are explicitly programmed. This thing LEARNED to micro.

>>	Anonymous Fri Jan 25 08:12:04 2019 No.10326560 >>10326519 But openai's dota2 bot didn't win

Anonymous Fri Jan 25 08:15:50 2019 No.10326563

>>10326054
>People have been trying for over 10 years to write bots that could play Starcraft
This is why you can't trust anyone on reddit lmao he doesn't even understand what AI is, the game shipped with fucking AI included that you could play against, what dos he even think he's talking about? He shills for deepmind but clearly has no grasp of the project goals.

>>	Anonymous Fri Jan 25 08:16:48 2019 No.10326567 >>10326556 That's the easiest shit to learn for an AI

>>	Anonymous Fri Jan 25 08:21:06 2019 No.10326576 >>10326567 So its easy for an AI to learn... what is your point?

>>	Anonymous Fri Jan 25 08:26:22 2019 No.10326588 >>10326560 It's probably one of the hardest games to get right for AI.

>>	Anonymous Fri Jan 25 08:37:12 2019 No.10326602 >>10326588 They used the same tactics as in this live game. The Brazilians split pushed on the base from multiple angles and open ai couldn't handle it. In this game mana dropped immortals into the base forcing the army back the same way.

>>	Anonymous Fri Jan 25 08:40:19 2019 No.10326608 >>10326588 >>10326602 Deep mind also won the same way openai won in the single lane 1v1 with just incredible mechanics

>>	Anonymous Fri Jan 25 08:41:34 2019 No.10326611 >>10322037 >>10323944 >>10323954 >>10323959 >>10323974 >>10323978 >>10323984 >>10324013 >>10324018 >>10324029 >>10324034 >>10324059 >>10324065 >>10324077 >>10324095 >>10324148 >>10324152 >>10324162 >>10324177 >>10324220 Samefag

>>	Anonymous Fri Jan 25 08:58:05 2019 No.10326638 >>10322077 that's because starcraft is a real game not a game for bitches

Anonymous Fri Jan 25 09:08:48 2019 No.10326659

Did they program it so that the AI has the same vision and control inputs that the human has?
Or does the AI "see" privileged information of all units within the field of vision every frame, can select any combination of units instantly, etc.
I don't find this that interesting unless the AI has to use a virtual mouse and keyboard and only sees at any given time what is on the screen and mini-map.

>>	Anonymous Fri Jan 25 09:10:10 2019 No.10326661 >>10326608 I guess we now know how humans can beat robots at warfare. Have to use unpredictable tactics where you split push in smaller units to cause damage behind lines.

>>	Anonymous Fri Jan 25 09:12:46 2019 No.10326667 >>10326659 The 10 games it won it had perfect information of everything on the map it had vision of and didn't need to move a camera (doesn't see into fog of war though)

>>	Anonymous Fri Jan 25 09:15:15 2019 No.10326673 >>10326661 That's in general how sc2 is played, but pvp os the worst matchup to see this

>>	Anonymous Fri Jan 25 09:24:02 2019 No.10326692 >>10326563 The game's AI is also argueably better at macro than Alpha. It does come down to non--human micro, which has nothing to do with intelligence.

>>	Anonymous Fri Jan 25 09:26:46 2019 No.10326695 >>10326659 This is actually what they went for a year ago and horribly failed, now they opted out to creating a hyped up microbot and call it a day. DeepMind really lost a lot of respect from me with this and the chess thing.

>>	Anonymous Fri Jan 25 09:28:45 2019 No.10326701 >>10326695 what happened with chess? they beat go so how could chess be a problem?

Anonymous Fri Jan 25 09:37:06 2019 No.10326717

>>10326701
They played against a dumbed down stockfish, won every game, declared Alpha the best chess AI in the world and didn't react to all the people calling bullshit.

AlphaZero is optimised for Go, they are just marketing it as "the AI that can learn anything in matters of weeks" because they want to sell it to governments.

Anonymous Fri Jan 25 09:38:05 2019 No.10326720

>>10326701
1. It seems that the hardware for AlphaZero vs Stockfish was uneven. The Stockfish developers saw it as unfair, not that Stockfish would've won on equal hardware, but that the results would've been different.
2. More speculative, but there are suspicions that DeepMind cherry-picked the games they published.
That said, it's likely that AlphaZero was in fact the strongest chess entity.

>>	Anonymous Fri Jan 25 09:40:44 2019 No.10326724 >>10326717 >>10326720 Yep Alpha won because it was A BETTER ENGINE. Not because it came up with new imaginative lines and whatever bulls hit they tried to sell us. Play Stockfish 10 again under TCEC conditions and see if it can win.

>>	Anonymous Fri Jan 25 09:46:09 2019 No.10326732 >>10326724 >A BETTER ENGINE Alpha had two advantages: Computing power, and an algorithm that could use them.

Anonymous Fri Jan 25 09:48:14 2019 No.10326734

>>10326724
You have to admit that AlphaZero had a really nice style. It definitely feels like "chess beyond the Stockfish era".
Let's just wait for Leela Chess Zero to surpass Stockfish. DeepMind are clearly a bunch of shady proprietary developers. Leela is based and freepilled

>>	Anonymous Fri Jan 25 09:54:29 2019 No.10326740 >>10326485 no, you're the real brainlet here. POMDPs have been applied to the room 'clearing' problem with pretty good success, at least in simulation.

>>	Anonymous Fri Jan 25 10:03:18 2019 No.10326755 >>10326740 >POMDPs have been applied to the room 'clearing' problem Blatantly incorrect use of terminology there, smart guy. POMDP is a class of problems, not a solution method.

>>	Anonymous Fri Jan 25 10:04:38 2019 No.10326756 >>10326724 >TCEC conditions AlphaZero runs on TPUs, making it ineligible for TCEC.

Anonymous Fri Jan 25 10:17:11 2019 No.10326782
File: 91 KB, 720x960, 1508539863341.jpg [View same] [iqdb] [saucenao] [google]

I just watched the entire 3 hours of footage.

I played Brood War (starcraft 1) about 8 hours a day for years in my youth. I have a couple of hundred hours of Starcraft 2 under my belt as well but never played it on a high level like brood war.

All-in-all the way AlphaStar played the game against TLO was very impressive and human-like. The AI took risks and gambled with things that paid off. The APM was reasonably low and they were mostly won through superior decision making and some luck instead of inhuman micro.

However against Mana and especially in Game 4 the AI played extremely unfair. The entire game was won by sheer inhuman APM at the most important confrontation it even peaked at around 1700 APM which were all individually perfectly coordinated in a way that would be biologically impossible for a human to accomplish. The actual strategy involved was actually on the human player's side and would have won under normal circumstances. This was plainly unfair and unimpressive.

Honestly I feel like this AI is meant to look like it's playing in a human way but when it feels "cornered" it resorts to inhuman micro to push for its true advantage over humans. The only reason it lost 1 round was because the human player tried to go for incomplete information to prevent the AI from doing these inhuman APM actions because the AI refrained from using them when it didn't know exactly what the enemy has behind the fog of war.

TL;DR: It's both impressive and disappointing at the same time. I feel like the AI only won 3 or 4 out of the 6 shown matches fairly which could potentially lead to a 3:3 stalemate between the AI and human players. Also the first player wasn't even a protos player so it should barely count as going against a pro. Against the actual pro the AI resorted to all sorts of inhuman plays to force a win in an otherwise lost match.

>>	Anonymous Fri Jan 25 10:51:07 2019 No.10326841 >Want it to develop strategy. >In an environment where micro is the winning move. Find something else to do.

>>	Anonymous Fri Jan 25 10:53:00 2019 No.10326842 Is it true that you are not smart until you beat a Korean in a Starcraft match?

>>	Anonymous Fri Jan 25 11:04:55 2019 No.10326863 >>10321959 >make min/max bot that does everything pre planned. >claim it's smart because of top secret neural network code. top kek, they've had these SC bots for decades.

Anonymous Fri Jan 25 11:07:32 2019 No.10326868

>>10325632
You know they did a bunch of games this year again right? Alpha crushed the newest version of stockfish too. And what made Alpha chess interesting was the way it played, not that it beat another engine. Its play looked more human than traditional chess bot's.

>>	Anonymous Fri Jan 25 11:09:15 2019 No.10326872 >>10322114 That's not true, in the posted match the AI is allowed to have an APM that is twice as high as the professional players' at the start of the match, which is the most important time.

>>	Anonymous Fri Jan 25 11:11:55 2019 No.10326879 >anything that AI can actually do isn't real AI keep moving them goalposts buckos

>>	Anonymous Fri Jan 25 11:13:37 2019 No.10326883 >>10324985 >in sc2 you have a built in hotkey for your entire fucking army 9 out of 10 times using this is a bad idea, and most higher level players rarely use it

>>	Anonymous Fri Jan 25 11:14:19 2019 No.10326886 >>10326841 Yeah Starcraft definitely isn't the game for this, there's some macro strategy but most of the meat of the game is micro which any old shitty hardcoded AI could do

>>	Anonymous Fri Jan 25 11:14:52 2019 No.10326888 >>10325009 >. in bw you need to move your screen to your barracks, individually select each one and queue a marine. this is arguably micro as well macro is your overall strategy

Anonymous Fri Jan 25 11:17:18 2019 No.10326894

>>10326886
there are different kinds of micro though. for example blinking back wounded stalkers is pretty trivial. However the kind of micro that alphastar was showing was beyond this.

The most striking example was when it was battling mana in the middle of the map. It split its army into three parts and was boxing mana's forces in. This clearly shows a kind of tactical strategy beyond just high APM micro.

Anonymous Fri Jan 25 11:20:46 2019 No.10326900

>>10326894
It was probably just running each unit to the nearest high ground at the time which naturally split the forces, not too crazy
When it actually came time for the AI to split its forces to deal with drop harass, it was just A moving its entire army in confusion

Anonymous Fri Jan 25 11:22:10 2019 No.10326903

>>10326782
>mostly won through superior decision making and some luck instead of inhuman micro.
What I found disappointing with the whole deal was that the engine didn't seem capable of any decision making other than basic tactical play(when to push, when to retreat), which is impressive together with it being able to play the game at all. But when it came to builds and what units to make, it didn't seem like it was making any decisions. After the agent had chosen its strategy during its learning, then that was it, no matter what its opponent did.

When it made its entire army of stalkers hopelessly chaise a dropship again and again into its main instead of being ready for it the next time, then I was really disillusioned with its intelligence.

This wasn't a big step beyond OpenAI's dota bot.

>>	Anonymous Fri Jan 25 11:30:12 2019 No.10326922 >>10326782 It looked better strategically because TLO played like a diamond/low masters player

Anonymous Fri Jan 25 11:30:17 2019 No.10326923

>>10326894
>The most striking example was when it was battling mana in the middle of the map. It split its army into three parts and was boxing mana's forces in. This clearly shows a kind of tactical strategy beyond just high APM micro.
Watch it again. Mana is the one that moves in between Alpha's forces and reinforcing units. It didn't split there purposefully for an advantage in the fight.

Anonymous Fri Jan 25 11:32:37 2019 No.10326930

>>10325926
Perfect micro bots in sc2 were made the same year the game was released and they didn't need neural networks to run. They want the AI they can not just outperform humans, but outsmart. In the game the pro won, he lured the AI's army out of position with a fairly weak harassment unit and simply marched his army in behind. The AI did nothing to counter the harassment unit and ended up stranded its army. A complete tactical missplay.

>>	Anonymous Fri Jan 25 11:36:00 2019 No.10326939 >>10326692 The in game AI on hardest difficulty gets more resources than the player and has perfect map vision. It literally cheats to make it more of a challenge for the players that have developed strategies

>>	Anonymous Fri Jan 25 11:57:18 2019 No.10326987 >>10321959 >only PvP >against TeamLoosing Ring me when it plays TvT against Flash.

Anonymous Fri Jan 25 11:57:49 2019 No.10326989

>>10326661
Except that is a tactic alphastar was using and had probably seen in every game it played. The problem was it never learned how to build specific counters or split it's army. Cannons or a handful of Phoenix (it already had the tech for both options) would have completely shut down the immortal drops. What did it do? Nothing, it just kept building the same shit and hoped MaNa would be stupid enough to fight it's stalker ball on alphastar's terms. MaNa just lured the stalker ball away from it's base and marched his army in. It didn't scout to find the army (several air units which are great at scouting were just sitting on the immortal drop keeping it scouted, it wasn't looking for mana's army). After MaNa's army walked in, it didn't even react. MaNa just walked all over it once he was in. Another thing is counter scouting. MaNa had an observer floating above alphastar's main ramp. No pro would have allowed that observer to sit there for more than a minute, they would have shut that scouting potential down. Hell, no player above plat rank would have let that observer sit there completely uncontested.

>>	Anonymous Fri Jan 25 11:58:19 2019 No.10326992 >>10326987 Not sure the AI could function without stalker teleporting

>>	Anonymous Fri Jan 25 12:01:47 2019 No.10327003 >practice starcraft for over 200 years >learn that pure stalker with perfect micro beats everything

Anonymous Fri Jan 25 12:04:54 2019 No.10327017

Here are some suggestion to the DeepMind team to make it fair.
Have the AI control physical mouse and keyboard through robotic arm.
Have input be a camera pointed at the screen.
It's unfair for it to see the whole map when players have to deal with the minimap.

>>	Anonymous Fri Jan 25 12:06:16 2019 No.10327021 >>10326992 I mean it's PvP and mechanical micro advantages with stalkers vs pros gives it a huge edge not that the results aren't impressive but deepmind basically tilted eveyrthing in their favor.

>>	Anonymous Fri Jan 25 12:06:48 2019 No.10327024 >>10327017 You greatly overestimate what neural networks are capable of.

>>	Anonymous Fri Jan 25 12:09:44 2019 No.10327030 >>10327024 How hard is it? Literally have a push button over every key of the keyboard, but limit it to what a human hand could do. have a horizontal 2 axis thing handling the mouse and 3 push things.

Anonymous Fri Jan 25 12:14:57 2019 No.10327042

>>10327030
desu you only really need to have an API for moving the mouse and the API takes into account various results from HCI research about the time it takes to move a mouse to hit something etc
literal mechanical shit would just be overkill unnecessary fluff

>>	Anonymous Fri Jan 25 12:15:56 2019 No.10327046 >>10326987 Forgot >Same fucking map every fucking time. This thing still has a long way to go.

>>	Anonymous Fri Jan 25 12:17:08 2019 No.10327047 >>10327042 But it's the hardest part about learning SC. Just fucking looking at the minimap. Deepmind doesn't have this problem.

>>	Anonymous Fri Jan 25 12:24:38 2019 No.10327074 >>10327047 well you can limit access to screen + sound + minimap without a camera e.g the "grab units" or "do stat of detected enemy unit" functions only offer lists of units that are on the screen

Anonymous Fri Jan 25 12:28:23 2019 No.10327077

>>10327042
I'd like to see if it could still do it with it, though.
I want this shit to have to deal with what we do.
You can't teleport your mouse pointer.
You can't click 100 times per second.
You can't see the whole map.
You have to click the minimap to see what that red dot is about.
Moving mouse takes time, and handling keyboard is an art of itself.
>>10327074
I guess, but if you go through so screen capture, it will have perfect vision of the whole screen, which isn't how human vision works.

>>	Anonymous Fri Jan 25 12:40:08 2019 No.10327101 File: 25 KB, 1280x967, 1280px-AcuityHumanEye.svg.png [View same] [iqdb] [saucenao] [google] >>10327077 Eyes are garbage.

>>	Anonymous Fri Jan 25 13:50:07 2019 No.10327296 >>10326542 >The point of this AI was to play like a human No it wasn't

>>	Anonymous Fri Jan 25 15:23:41 2019 No.10327506 >>10326989 The question is why the ai acted that way. Why didn't it react properly?

>>	Anonymous Fri Jan 25 15:29:59 2019 No.10327531 >>10327506 My guess is that the training is not exploring all these small details well so it just doesn't know

Anonymous Fri Jan 25 15:51:01 2019 No.10327594

>>10323127
alphaZero was general enough to solve three different games (chess, shogi, go) with the same basic algorithms, but the network architecture itself was customized to the board and move topologies of each of the games.

Also it was using a particular search strategy (MCTS) instead of coming up with one on its own. We are still a long, long way from deep learning inventing search techniques like alpha-beta or MCTS all by themselves.

Anonymous Fri Jan 25 23:13:55 2019 No.10328838

>>10327077
>I'd like to see if it could still do it with it, though.
It won't be able to, it can't even win reliably with full API access. Hell AlphaGo wouldn't be able to play either if it had to move the stones physically on its own.

>>10327506
Because from what I can tell they're just using a glorified policy gradient, it's basically just brute force. Start with random actions and if you get good results tilt the distribution towards that.
It doesn't reason about anything, and if it encounters an unfamiliar state that it didn't train on for 50 years of game time it will fail miserably.

Advanced search
Text to find
Subject [?]Search by post subject. Leave empty for any.
Username [?]Search for user name. Leave empty for any user name.
Tripcode [?]Search for tripcode. Leave empty for any.
Email [?]Search by email. Leave empty for any.
Filename [?]Search by image filename. Leave empty for any.
From Date [?]Enter what date to start searching from. Format is YYYY-MM-DD
To Date [?]Enter what date to start searching until. Format is YYYY-MM-DD
Image hash
Search in	All Posts OPs Only
Deleted posts	Show all posts Show only deleted posts Only show non-deleted posts
Internal posts	Show all posts Show only internal posts Show only archived posts
Order	New posts first Old posts first
Capcode	All Posts Only by Users Only by Mods Only by Admins Only by Developers
Results	Posts Threads
Action	[ Simple ]

/sci/ - Science & Math