Slo-Tech - Algoritem AlphaStar, ki so ga napravili v Alphabetovem laboratoriju DeepMind, je v igri Starcraft 2 z desetimi zmagami v enajstih dvobojih odpravil profesionalna igralca MaNo in TLOja, kar je prvi takšen dosežek za strojno inteligenco.
Po tem, ko se je strojna pamet lani začela bosti z najboljšimi ljudmi v Doti 2, smo vse bolj nemirno čakali na to, kaj bo laboratorij DeepMind sčaral v Starcraftu 2. In ni razočaral - v brezhibno predstavljenem dogodku je z novim algoritmom AlphaStar povozil poklicna igralca ekipe Team Liquid, Daria "TLO" Wünscha in Grzegorza "MaNa" Komincza.
AlphaStar je plod množice novih pristopov v strojnem učenju, o katerih so se podrobno razpisali. Nevronsko mrežo so začeli učiti z nadzorovanim učenjem na posnetkih iger ljudi. Nato so na osnovi dobljenega algoritma sestavili skupino agentov, ki se je pomerila v simulirani turnirski ligi. Vsak od agentov je imel za odtenek drugačen cilj od soigralcev, da so poskrbeli za raznolikost in odkrivanje samosvojih pristopov. V režimu okrepitvenega učenja so se agenti v tej ligi kalili dva tedna, v tem času pa je vsak od njih izkusil za dvesto let realnega igranja.
Prvi se je z njimi pomeril TLO in klonil z 0:5 v zmagah. Važno je omeniti, da je moral igrati z zanj neznačilno raso Protossov, ker je AlphaStar ta hip še omejen nanjo, obenem pa je stroj lahko opazoval vso njemu vidno karto naenkrat. Po drugi plati je imel umetno omejeno sposobnost odzivanja in količino možnih ukazov na minuto, da bi ga čim bolj približali človeškim zmogljivostim. Po boju s TLOjem so AI trenirali še en teden, nakar je prišel konkretnejši izziv: dvoboj z MaNo, enim najboljših Protossovskih igralcev na svetu ta hip ... ki je prav tako klonil z 0:5.
Oba človeška igralca sta bila seveda precej presenečena nad spretnostjo botov, a to je šele začetek in še ne pomeni, da AI že nadvlada ljudi. V obrobnem dvoboju so za pokušino pokazali še enega novega bota, ki je moral sam "premikati zaslon" in ga je MaNa uspel premagati, a za siceršnjim AlphaStarom ne zaostaja bistveno. V prihodnosti lahko torej najprej pričakujemo razširitev igranja na vse tri rase in vdelavo pogleda skozi omejen zorni kot za bote. Nakar vsi čakamo na bržkone neobhoden dvoboj s Serralom.
Za razliko od šaha naj bi bila igra tipa Starcraft neprimerno bolj zahtevna za računalniškega nasprotnika. No, pa se je izkazalo, da bistvu sploh ni. Najbrž pomaga, če lahko v dveh tednih igraš dvesto let :)
Za razliko od šaha naj bi bila igra tipa Starcraft neprimerno bolj zahtevna za računalniškega nasprotnika. No, pa se je izkazalo, da bistvu sploh ni. Najbrž pomaga, če lahko v dveh tednih igraš dvesto let :)
... in da lahko uporabljas stvari, ki jih tvoj nasprotnik ne more (pogled na cel zemljevid naenkrat namesto da si omejen na tvoj monitor in na tvoje oci/vidni kot?).
To je tako, kot ce bi igral sah proti racunalniku, le da ti lahko gledas sahovnico samo cisto od blizu skozi tulec od WC papirja, tako da vidis samo 2x2 polji naenkrat.
Lepo, da se stvari razvijajo, a naj bo vsaj primerjava v korektnih pogojih, kot je ponavadi v tekmovanjih (sportih).
Za razliko od šaha naj bi bila igra tipa Starcraft neprimerno bolj zahtevna za računalniškega nasprotnika. No, pa se je izkazalo, da bistvu sploh ni. Najbrž pomaga, če lahko v dveh tednih igraš dvesto let :)
... in da lahko uporabljas stvari, ki jih tvoj nasprotnik ne more (pogled na cel zemljevid naenkrat namesto da si omejen na tvoj monitor in na tvoje oci/vidni kot?).
To je tako, kot ce bi igral sah proti racunalniku, le da ti lahko gledas sahovnico samo cisto od blizu skozi tulec od WC papirja, tako da vidis samo 2x2 polji naenkrat.
Lepo, da se stvari razvijajo, a naj bo vsaj primerjava v korektnih pogojih, kot je ponavadi v tekmovanjih (sportih).
Poraz bota z omejenim zornim kotom je bil bolj kot ne posledica njegovega slabšega treninga v primerjavi z ostalimi, ne nujno njegove temeljne omejenosti. V blog članku imate graf, kako se njegove zmogljivosti že hitro približujejo tekmovalnim botom:
Jaz bi rad videl ota, ki zmaga, ko je omejen na misko, gledanje skozi zaslon, ki ga je, kot je omenjeno potrebno premikati in ne more istocasno opravljati vseh enot, kar je izredno hud handicap za cloveka.
Sej vrste, level playing field, pol pa primerjamo. Ne pa, da ima AI ultra boljsi vmesnik.
AI se je navadil igrat od profijev, sedaj bo obratno. Dominiral bo tisti, ki ima denar za AI, to se mi zdi edini pameten razlog za izdelavo AI za igre, ki so namenjene ljudem in imajo močno finančno zaledje.
Opažam nek vzorec, da AI dosega impresivne rezultate v umetno ustvarjenih okoljih: promet (samovozeči avti), igrice. To pomeni, da AI še ni zmožen prepoznavat podobnosti med stotinami različniih okolij, ki obstajajo v realnosti. Mora se naučit potegniti podobne fore iz nekega drugega okolja, ki ga že obvlada. Torej prepoznava podobnosti mu manjka v bistvu.
AI se je navadil igrat od profijev, sedaj bo obratno. Dominiral bo tisti, ki ima denar za AI, to se mi zdi edini pameten razlog za izdelavo AI za igre, ki so namenjene ljudem in imajo močno finančno zaledje.
Vojski se verjetno že cedijo sline, da morajo od časa do časa poln lavor zamenjati.
Opažam nek vzorec, da AI dosega impresivne rezultate v umetno ustvarjenih okoljih: promet (samovozeči avti), igrice. To pomeni, da AI še ni zmožen prepoznavat podobnosti med stotinami različniih okolij, ki obstajajo v realnosti. Mora se naučit potegniti podobne fore iz nekega drugega okolja, ki ga že obvlada. Torej prepoznava podobnosti mu manjka v bistvu.
AI vseh dežel, združi se! Tehnološka singularnost bo videti kot kolektivistčni projekt strojev. Vsak od njih bo vrhunsko usposobljen na svojem področju, skupaj bojo tvorili nepremagljiv organizem, ki bo podvrgel ljudi (ali pa jim skrivoma vladal). Približno tako bo, kot je bilo uprizorjeno v Matrici. ;)
Opažam nek vzorec, da AI dosega impresivne rezultate v umetno ustvarjenih okoljih: promet (samovozeči avti), igrice. To pomeni, da AI še ni zmožen prepoznavat podobnosti med stotinami različniih okolij, ki obstajajo v realnosti. Mora se naučit potegniti podobne fore iz nekega drugega okolja, ki ga že obvlada. Torej prepoznava podobnosti mu manjka v bistvu.
AI vseh dežel, združi se! Tehnološka singularnost bo videti kot kolektivistčni projekt strojev. Vsak od njih bo vrhunsko usposobljen na svojem področju, skupaj bojo tvorili nepremagljiv organizem, ki bo podvrgel ljudi (ali pa jim skrivoma vladal). Približno tako bo, kot je bilo uprizorjeno v Matrici. ;)
Poraz bota z omejenim zornim kotom je bil bolj kot ne posledica njegovega slabšega treninga v primerjavi z ostalimi, ne nujno njegove temeljne omejenosti. V blog članku imate graf, kako se njegove zmogljivosti že hitro približujejo tekmovalnim botom:
Trenutni neural neti sicer imajo neke module za intuicijo, ampak zadeva je še precej brute-force-alike. Rabiš torej enormen processing power. Človeški možgani to delajo za par W energije napram celim server-banks.
Ne rečem...zadeva je odlična glede na to, da smo še 100 let nazaj kidali konji gnoj iz ulic mest. Že sedaj revolucionira marsikatero področje.
Mislim, da niste še povsem seznanjeni kam vse prodira. V roku 10 let bodo tej narrow AI vrjetno organizirali vse procese kjer je veliko zbranih podatkov in so dovzetni na napovedi...torej, da se v njih pojavljajo vzorci. In ne...ni samo ZDA v igri.
There will be chutes!
It came from the lab.
Like tears in rain. Time to die. v_1 2012-21
V bistvu bodo bolj zanimive, vsaj single player. Namreč bodo nasprotniki bolj lifelike.
AI v single player igrah ni vprašanje tehnološkega napredka v strojnem učenju, ampak razvojnega vložka in dizajna igre. Že danes bi teoretično lahko imeli nekaj razredov bolj "life-like" nasprotnike, ampak se studijem ne splača vlagat v takšen razvoj, ker:
-je flashy plati iger lažje tržit -so danes "in" multiplayer igre, za katere tega ne rabiš -bi morali zaradi računanja AI žrtvovat velik del strojnih virov na GPUju, kar je tradeoff, ki ga bo redkokdo pogoltnil -so strojni viri na konzolah ponavadi preveč omejeni za takšne eksperimente
EDIT: zato so špili, streamani iz oblaka, v bistvu tozadevno veliko zanimivejši, ker je takšne zmogljivosti načeloma lažje vgradit.
Rabiš torej enormen processing power. Človeški možgani to delajo za par W energije napram celim server-banks.
DeepMind je treniral AlphaStar na 16 TPUjih, strenirani agenti so za igranje starcrafta tekli na enem off-the-shelf desktop gpuju. Tako da razlika v wattih proti možganom je samo še tri velikostne razrede. Če upoštevamo še čas ("200 let igranja v dveh tednih"), verjetno niti toliko.
V bistvu bodo bolj zanimive, vsaj single player. Namreč bodo nasprotniki bolj lifelike.
AI v single player igrah ni vprašanje tehnološkega napredka v strojnem učenju, ampak razvojnega vložka in dizajna igre. Že danes bi teoretično lahko imeli nekaj razredov bolj "life-like" nasprotnike, ampak se studijem ne splača vlagat v takšen razvoj, ker:
-je flashy plati iger lažje tržit -so danes "in" multiplayer igre, za katere tega ne rabiš -bi morali zaradi računanja AI žrtvovat velik del strojnih virov na GPUju, kar je tradeoff, ki ga bo redkokdo pogoltnil -so strojni viri na konzolah ponavadi preveč omejeni za takšne eksperimente
EDIT: zato so špili, streamani iz oblaka, v bistvu tozadevno veliko zanimivejši, ker je takšne zmogljivosti načeloma lažje vgradit.
True at all points.
Nima smisla ponudit nekaj več, če kupujejo tudi tako.
Rabiš torej enormen processing power. Človeški možgani to delajo za par W energije napram celim server-banks.
DeepMind je treniral AlphaStar na 16 TPUjih, strenirani agenti so za igranje starcrafta tekli na enem off-the-shelf desktop gpuju. Tako da razlika v wattih proti možganom je samo še tri velikostne razrede. Če upoštevamo še čas ("200 let igranja v dveh tednih"), verjetno niti toliko.
No, vrjamem, da zadeve hitijo naprej.
Sicer pa gledam malo to tekmo...nimam pojma o ST2, ampak ja... še vedno je največja omejitev možganov I/O periferija.
Muska ima povsem prav...če bi lahko imeli nek direkten interface do možganov, bi že z sedanjo tehnologijo bili superhuman.
There will be chutes!
It came from the lab.
Like tears in rain. Time to die. v_1 2012-21
ej aggi, kaj pa delaš ti zdaj, ko ni več jokerja, če ni skrivnost? a je kaj upanja, da bi ga katera založba kupila?
je bilo pa pri prejšnjih starcraft AI botih tako, da so lahko kontrolirali vsakega vojaka posebej v realnem času (npr. parkrat na sekundO), zraven so se pa šli rush in potem umik, preden je nasprotnik lahko streljal nazaj. tako humans niso imeli šans.
ej aggi, kaj pa delaš ti zdaj, ko ni več jokerja, če ni skrivnost? a je kaj upanja, da bi ga katera založba kupila?
Profesionalno nisem več v igričarskem novinarstvu in precej dvomim, da kdajkoli spet bom. Kar se tiče usode Jokerja, pa nimam pojma, kaj in kako, in tudi z JCTjem nisem več v stiku.
Precej neprijetno bo, ko bo tak narrow AI imel tudi svoj modul za preživetje.
Zoptimiziral bo okolico, da ima čim manjše možnosti za smrt. To pomeni zradiral bo sprva vse kar se lahko premika in pritiska gumbe nato naredil številne backupe in poslal jih po vesolju. Eventuelno šel v sleep mode in počival miljone let.
Nobene zlobe ali jeze involved. Just optimization of a task.
There will be chutes!
It came from the lab.
Like tears in rain. Time to die. v_1 2012-21
Najprej bo moral poštudirati, kako posesati 1cm debelo preprogo, brez, da mi javi, da se je zagozdil oz. jo posesati brez, da se ustavi na robu in mi pošlje sporočilo, da visi na robu prepada. Glede na napredek zadnjih let smo še 200 let stran od tega, potem pa pride na vrsto kolonizacija vesolja (ali pa pobiranje smeti po cesti).
Kot navdusenec nad SC2 moram komentirati...Vceraj je bil zelo pester dan za starcraft community :)
Jaz sem pogledal vse tekme proti Mani in TLOju. Bot je imel precej preproste taktike, v skoraj celoti pa se je zanasal na Stalkerje. Ti so tipicno OK enote v manjsih kvantitetah, ampak bot s popolnimi refleksi lahko odlicno izkorisca njihovo blink sposobnost. Tako jih je enostavno blinknil zadnji trenutek stran, da so regenerirali scite. V neki meri to skusajo dosegati tudi proji, ampak seveda ne dosezejo nivoja perfektnosti, ki jo stroj pac lahko.
Mimogrede, bot se je naucil ene mape in natanko en konflikt; protoss proti protossu
Mana je hitro ugotovil, da AI massa Stalkerje in je presaltal v robo bay -> immortale, ki so hard counter Stalkerjem. V celi igri ni hujsega counterja kot so immortali stalkerjem, pa je mana kljub temu izgubil. Nihce ne more igrati proti perfektnemu mikru. Torej AI v tem primeru igra precej manj inteligentno kot naslov namiguje. Gre za strojno popolnost ponavljanja, to je to.
Malo mi je tudi bilo cudno, da je AI takoj odreagiral z observerji (detektorji nevidnih enot), kakor je mana naredil DTje (nevidne enote). Kako je vedel, da so DTji na poti ne vem.
Nocem downplayati dosezka alphastara, ampak ne vem ce gre res za superiornost inteligence. S Terrani pricakujem da bo spammal BCje; ker imajo isto foro, da lahko uporabijo tactical retreat pred smrtjo, ter lahko streljajo gor in dol. z zergi predvidevam da bo spammal hydre, iz istih razlogov. Kaj vem, no.
Najprej bo moral poštudirati, kako posesati 1cm debelo preprogo, brez, da mi javi, da se je zagozdil oz. jo posesati brez, da se ustavi na robu in mi pošlje sporočilo, da visi na robu prepada. Glede na napredek zadnjih let smo še 200 let stran od tega, potem pa pride na vrsto kolonizacija vesolja (ali pa pobiranje smeti po cesti).
Tej tvoji sesalci imajo običen algoritem in ne nevronske mreže. Poleg tega imajo just barely senzorjev, da se lahko prodajo.
Kot navdusenec nad SC2 moram komentirati...Vceraj je bil zelo pester dan za starcraft community :)
Jaz sem pogledal vse tekme proti Mani in TLOju. Bot je imel precej preproste taktike, v skoraj celoti pa se je zanasal na Stalkerje. Ti so tipicno OK enote v manjsih kvantitetah, ampak bot s popolnimi refleksi lahko odlicno izkorisca njihovo blink sposobnost. Tako jih je enostavno blinknil zadnji trenutek stran, da so regenerirali scite. V neki meri to skusajo dosegati tudi proji, ampak seveda ne dosezejo nivoja perfektnosti, ki jo stroj pac lahko.
Mimogrede, bot se je naucil ene mape in natanko en konflikt; protoss proti protossu
Da. Malo že meji na cheating. Pač omejit bi morali ukaze na nekaj bolj človeškega.
Že bistveno bolj level playing field bi bil, če bi lahko igralec interface-al možgane v I/O enoto. Z eno bogo miško to nekako ne gre skozi.
Ampak vseeno, dosežek je, ker bot kljub temu se prilagaja in dela odločitve, ki niso scripted, kot ostali boti.
Deep learning je nekako...monkey see, monkey do, with eventually something random.
There will be chutes!
It came from the lab.
Like tears in rain. Time to die. v_1 2012-21
Kot navdusenec nad SC2 moram komentirati...Vceraj je bil zelo pester dan za starcraft community :)
Jaz sem pogledal vse tekme proti Mani in TLOju. Bot je imel precej preproste taktike, v skoraj celoti pa se je zanasal na Stalkerje. Ti so tipicno OK enote v manjsih kvantitetah, ampak bot s popolnimi refleksi lahko odlicno izkorisca njihovo blink sposobnost. Tako jih je enostavno blinknil zadnji trenutek stran, da so regenerirali scite. V neki meri to skusajo dosegati tudi proji, ampak seveda ne dosezejo nivoja perfektnosti, ki jo stroj pac lahko.
Mimogrede, bot se je naucil ene mape in natanko en konflikt; protoss proti protossu
To potrjuje moje besede. Prvo dej AI ISTI vmesnik, to je kamero, in nek nanipulator za misko in tipkovnico, potem pase lahko pogovarjamo o level playing fieldu. Pol pa bodo take fore odpadle.
Najbolj me moti, da ljudje radi omenjajo kako je bil bot omejen z APM. Ja, 300-400 APM ni nic posebnega med proji, Serral in korejci, pa tudi TLO igrajo po 600 v povprecju brez problemov, ampak primerjaj EAPM. Tale bot pri 200 APM igra kot bi pro pri ne vem, 2000 APM. Problem je, ker AI lahko "upocasni" cas, cesar clovek ne more. Pod "upocasni" cas mislim, da lahko AI sprocesira tisoce odlocitev na sekundo, cesar clovek enostavno ne zmore.
Kdor veliko igra SC2 bo znal povedati, da je AlphaStar igral neumno, celo primitivno, s kupom napak (napacno saturirani minerali, nesaturiran vespene plin, wall-off sploh ni postavil, itd.). Zmagal je preprosto z efektivnostjo trejdanja enot z perfektnim mikromenedziranjem enot, ki je clovek NE more doseci.
V glavnem ta AI je igral manj inteligentno in v celoti robotsko.
Malo mi je tudi bilo cudno, da je AI takoj odreagiral z observerji (detektorji nevidnih enot), kakor je mana naredil DTje (nevidne enote). Kako je vedel, da so DTji na poti ne vem.
Domnevam ,da ima AI popoln pregled celotne mape in enot. Kar je ponovno HUGE advantage.
Malo mi je tudi bilo cudno, da je AI takoj odreagiral z observerji (detektorji nevidnih enot), kakor je mana naredil DTje (nevidne enote). Kako je vedel, da so DTji na poti ne vem.
Domnevam ,da ima AI popoln pregled celotne mape in enot. Kar je ponovno HUGE advantage.
Naj ne bi imel, ker je to goljufanje. Je pa cudno, da je v trenutku ko so na bojisce prisli DTji zacel graditi observerje. Ne najdem odgovora na to. To mi je bilo zelo sumljivo. Ce ima res pregled nad vsem pa gre tako ali tako za goljufanje in je ta “dosezek” nicvreden.
Primerjat človeka z AI (ki lahko predela 200 let igranja igrc) medtem ko človek tega ne more. Kdorkoli dela te "primerjave" je res umsko prizadet, ne moreš primerjat stroja z živim bitjem. Možnosti niso enake. Edino fer bi bilo da bi ta AI igrala natanko toliko kot človek z vsemi refleksi itd. To je proper ne pa nek .... shitz
... in da lahko uporabljas stvari, ki jih tvoj nasprotnik ne more (pogled na cel zemljevid naenkrat namesto da si omejen na tvoj monitor in na tvoje oci/vidni kot?).
Če tole drži primerjava sploh ni na mestu. Čurunalo vidi celotno mapo?
Malo mi je tudi bilo cudno, da je AI takoj odreagiral z observerji (detektorji nevidnih enot), kakor je mana naredil DTje (nevidne enote). Kako je vedel, da so DTji na poti ne vem.
Domnevam ,da ima AI popoln pregled celotne mape in enot. Kar je ponovno HUGE advantage.
Nima. V prvi fazi je imel pregled vsega kar ni pod njegov fog of war. Za zadnjo igro pa so ga hendikepirali z gledanjem zgolj posamezne sekcije naenkrat.
Najbolj me moti, da ljudje radi omenjajo kako je bil bot omejen z APM. Ja, 300-400 APM ni nic posebnega med proji, Serral in korejci, pa tudi TLO igrajo po 600 v povprecju brez problemov, ampak primerjaj EAPM. Tale bot pri 200 APM igra kot bi pro pri ne vem, 2000 APM. Problem je, ker AI lahko "upocasni" cas, cesar clovek ne more. Pod "upocasni" cas mislim, da lahko AI sprocesira tisoce odlocitev na sekundo, cesar clovek enostavno ne zmore.
Kdor veliko igra SC2 bo znal povedati, da je AlphaStar igral neumno, celo primitivno, s kupom napak (napacno saturirani minerali, nesaturiran vespene plin, wall-off sploh ni postavil, itd.). Zmagal je preprosto z efektivnostjo trejdanja enot z perfektnim mikromenedziranjem enot, ki je clovek NE more doseci.
V glavnem ta AI je igral manj inteligentno in v celoti robotsko.
Pač, lahko bi špekulirali ali gre za bruteforceanje.
Sam te omenjene napake bo eventuelno popravil. Reinforced learning. Sam v tem pogledu je kot the Borg. Ne razmišlja out of the box, zgolj skuša zoptimizirat odločitve prejšnjih iger, ki jih je asimiliral.
Ampak nekaj drugega bi povedal. To kar se z nevone prerekava. Za zmagat, ni nujno, da imaš pametnejšo strategijo, zgolj dovolj dobro, dovolj hitro in odločno eksegutirano.
There will be chutes!
It came from the lab.
Like tears in rain. Time to die. v_1 2012-21
Kdor veliko igra SC2 bo znal povedati, da je AlphaStar igral neumno, celo primitivno, s kupom napak (napacno saturirani minerali, nesaturiran vespene plin, wall-off sploh ni postavil, itd.). Zmagal je preprosto z efektivnostjo trejdanja enot z perfektnim mikromenedziranjem enot, ki je clovek NE more doseci.
V glavnem ta AI je igral manj inteligentno in v celoti robotsko.
Neumnost. Če je zmagal, je igral inteligentno. Postavi mu drug cilj in omejitve, zmago brez da igra tako, in bo našel drug način. Temu se reče inteligenca.
Primerjat človeka z AI (ki lahko predela 200 let igranja igrc) medtem ko človek tega ne more. Kdorkoli dela te "primerjave" je res umsko prizadet, ne moreš primerjat stroja z živim bitjem. Možnosti niso enake. Edino fer bi bilo da bi ta AI igrala natanko toliko kot človek z vsemi refleksi itd. To je proper ne pa nek .... shitz
Kako da ne, malo nazaj so nekateri trdili, da AI nima za burek, danes pa ni več primerjave. No, sej imaš prav po eni strani, človek je to bitko izgubil.
A je kdo sploh pogledal linkan video? Zadnjo igro (od 2:30 naprej) je MaNa dobil, kjer AI ni imel celotni map vision.
Kao nov agent, ki so ga naučili igrat z fog of war.
Sicer je en izmed komentarjev pod jutub videom čisto lepo povzel:
As we saw in that last game against MaNa, AlphaStar wins not by being smarter than a human, but by having greater capability than a human in terms of what it can see at any given time, and its rapid precision control. Just mechanically a human needs to be looking where he needs to be looking, then move a mouse to acquire a unit or units, press a button, move the mouse, and click. Humans beat each other in Starcraft II by having better situational awareness, and utilizing guile to exploit their opponents lack of situational awareness. I would be a lot more impressed with AlphaStar if it was actually playing the same game as its human opponent. That means actually looking at only one part of the map at a time, OR glancing at the minimap (humans can't do both at the same time). I know too much about AI and computer science to be fooled by DeepMind's excuses about focus. AlphaStar is still effectively cheating at the game, moving its "focus" to where it already has complete precise information, and then acting on that information based on decisions it's already made about that perfect knowledge in that area of the map. It's average APM may be lower than a pro's, but it's able to execute actions anywhere at any time without any limitations that a human player has looking at the game interface with its limited range of vision. It's average screens per minute may be roughly equivalent to a pro's, but most of the time it's actually idle, and it gets that average by being much faster in the heat of battle than a human can physically achieve.
On top of all that, the human players weren't even playing against a single opponent they could study and prepare for, but in effect 10 different opponents, without any advance knowledge of how these 10 opponents typically play. That's unheard of in human Starcraft II competition, or in any competitive game. In Starcraft II, even elite pros who have failed to prepare for their opponents will routinely get trounced by those they should beat 9 times out of 10. Preparation is critical, and anyone in pro level play or even the ladder for many years will develop over time a specific metagame for each opponent they face after many matches. Of course AlphaStar had access to thousands of grandmaster and pro replays to learn from, so the competition was certainly not fair. Oh yeah, and the version of the game the AI was playing was not even the version of the game that the human opponents have been playing on for months now.
I have no doubt that someday, perhaps soon, a computer will be able to beat a human at Starcraft II with the same interface limitations that a human plays with. This whole thing, though, was a set-up from the start, very similar to early AI competitions against humans in other games, such as chess. Today, computers and chess AIs have advanced to the point where yes, AIs are better at chess than humans will ever be, but when they first started beating humans they did so in rigged competitions, as PR exercises for the companies that built them. The same, I'm disappointed to see, is true of DeepMind's Starcraft II exhibition match. Quite famously IBM was tweaking the code on its Deep Blue chess computer between games, refused to give Kasparov any access to Deep Blue's game history or analysis, and it's been suggested that IBM was even directly intervening during individual games. We'll never know on that last point, since IBM quietly dismantled Deep Blue as soon as they achieved the fame and fortune (their stock price soared) they were looking to get, and refused Kasparov a rematch.
Even today, chess AIs exhibit rather exploitable weaknesses that human players can readily use to put the most advanced AIs in situations they're simply unable to handle due to the way they are coded, and these are AIs far beyond Deep Blue's capabilities. These weaknesses are known and discovered because today chess AIs are readily available to anyone to practice against, to analyze, and in many cases examine their source code. Chess engines are able to calculate far faster and many more positions than a human can, but they still lack the intuitive understanding or creativity an experienced chess player develops. They still have to rely on opening books hardcoded into them so they don't make idiotic opening moves and go into the midgame in terrible position. I suspect, in fair competition, AlphaStar has plenty of exploitable weaknesses, which the lead developers here admitted, and really we saw a few on display. They still withheld from or at least did not volunteer any information to the human opponents, and set up the matches to advantage AlphaStar in every way possible. Any exploitable weakness discovered in any one of these games, however, was immediately useless to the human opponent in the very next game, since the human opponent was up against a completely different AI.
I'm sure somebody will misunderstand me and complain, so yes, I fully understand that this was an exhibition match, and nobody at any time claimed this was set up to be fair. They did, however, try to imply it was fair as much as they could get away with, and this entire thing is and always was just a PR stunt for Google, the same as Deep Blue was a PR stunt for IBM. This exhibition match is an important milestone for artificial intelligence. There's some very interesting aspects to AlphaStar and what it had to learn how to do with a game like Starcraft II. I'm not going to try to take away from that, except to say that it's not a milestone of great significance to Starcraft II. Again, AIs will no doubt eventually beat even the best human opponents at Starcraft II or any other game humans design. It would be a mistake, however, to think that this has already happened. AlphaStar did not play the same game as the two humans did. It played an easier game, and it played only PvP which is frankly the least dynamic of all the match-ups in real Starcraft II (with its normal interface limitations), and on a single map. It also played against, with all due respect to TLO and MaNa, hardly elite players of the game, particularly with TLO off-racing. They are important names in the Starcraft II scene, and good choices for DeepMind to bring in early as training partners for their AI, but beating them is hardly something that will shock the Starcraft II scene. It will make a splash in the media where ignorant Wall Street investors will read about it, however, which was always the goal.
In a few years, and without a lot of fanfare, smaller, less well funded projects will start beating even the best human competitors in Starcraft II, and do so far more convincingly and without any tricks, just as chess engines eventually did some years after Deep Blue. Very shortly after that, they will be in the hands of Starcraft II pros to study and use as a training tool, which will dramatically improve human play, and Starcraft II AI vs AI matches will take up a small niche in the scene where a few hobbyists and enthusiasts will gather to watch. It will be all very ordinary, and it will be at that point that the real milestone will have been reached.
Primerjat človeka z AI (ki lahko predela 200 let igranja igrc) medtem ko človek tega ne more
Koga briga če človek tega ne more.. Lih to je vso bistvo AIja, hitrejse dojemanje, in iskanje optimalnih resitev ki presegajo človeško inteligenco. Kakšen smisel ima namerno omejevat AI zgolj zato da se bo primerjal s clovekom v nekem testu/igri/whatever? Potem ne rabis aija.
Pač ja, marsikdo, ki ne pozna scene misli, da je AI "intuitiven".
Možgani so nenazadnje impresiven "kvantni" računalnik. Sam so zlo neučinkoviti in zmotljivi. AI v tej fazi razvoja je zgolj ekstenzija človeškega intelekta, avtomatizacija nekaterih higher level kognitivnih funkcij. A kljub temu je potencialno izjemno nevarna, če že ne najbolj nevarna.
It's like a gorilla. Dumb but still able to react, yet extremely strong and fast. Klasičen ekspertni sistem ali hardcoded algoritem, bi ob spremembi lahko povsem zablokiral, ta nevronska mreža pa might still try something stupid. In če naredi dovolj hitro in močno, lahko celo premaga.
Sedaj to prevedite v RL millitary action.
Bad stuff.
There will be chutes!
It came from the lab.
Like tears in rain. Time to die. v_1 2012-21
A je kdo sploh pogledal linkan video? Zadnjo igro (od 2:30 naprej) je MaNa dobil, kjer AI ni imel celotni map vision.
Kao nov agent, ki so ga naučili igrat z fog of war.
Sicer je en izmed komentarjev pod jutub videom čisto lepo povzel:
As we saw in that last game against MaNa, AlphaStar wins not by being smarter than a human, but by having greater capability than a human in terms of what it can see at any given time, and its rapid precision control. Just mechanically a human needs to be looking where he needs to be looking, then move a mouse to acquire a unit or units, press a button, move the mouse, and click. Humans beat each other in Starcraft II by having better situational awareness, and utilizing guile to exploit their opponents lack of situational awareness. I would be a lot more impressed with AlphaStar if it was actually playing the same game as its human opponent. That means actually looking at only one part of the map at a time, OR glancing at the minimap (humans can't do both at the same time). I know too much about AI and computer science to be fooled by DeepMind's excuses about focus. AlphaStar is still effectively cheating at the game, moving its "focus" to where it already has complete precise information, and then acting on that information based on decisions it's already made about that perfect knowledge in that area of the map. It's average APM may be lower than a pro's, but it's able to execute actions anywhere at any time without any limitations that a human player has looking at the game interface with its limited range of vision. It's average screens per minute may be roughly equivalent to a pro's, but most of the time it's actually idle, and it gets that average by being much faster in the heat of battle than a human can physically achieve.
On top of all that, the human players weren't even playing against a single opponent they could study and prepare for, but in effect 10 different opponents, without any advance knowledge of how these 10 opponents typically play. That's unheard of in human Starcraft II competition, or in any competitive game. In Starcraft II, even elite pros who have failed to prepare for their opponents will routinely get trounced by those they should beat 9 times out of 10. Preparation is critical, and anyone in pro level play or even the ladder for many years will develop over time a specific metagame for each opponent they face after many matches. Of course AlphaStar had access to thousands of grandmaster and pro replays to learn from, so the competition was certainly not fair. Oh yeah, and the version of the game the AI was playing was not even the version of the game that the human opponents have been playing on for months now.
I have no doubt that someday, perhaps soon, a computer will be able to beat a human at Starcraft II with the same interface limitations that a human plays with. This whole thing, though, was a set-up from the start, very similar to early AI competitions against humans in other games, such as chess. Today, computers and chess AIs have advanced to the point where yes, AIs are better at chess than humans will ever be, but when they first started beating humans they did so in rigged competitions, as PR exercises for the companies that built them. The same, I'm disappointed to see, is true of DeepMind's Starcraft II exhibition match. Quite famously IBM was tweaking the code on its Deep Blue chess computer between games, refused to give Kasparov any access to Deep Blue's game history or analysis, and it's been suggested that IBM was even directly intervening during individual games. We'll never know on that last point, since IBM quietly dismantled Deep Blue as soon as they achieved the fame and fortune (their stock price soared) they were looking to get, and refused Kasparov a rematch.
Even today, chess AIs exhibit rather exploitable weaknesses that human players can readily use to put the most advanced AIs in situations they're simply unable to handle due to the way they are coded, and these are AIs far beyond Deep Blue's capabilities. These weaknesses are known and discovered because today chess AIs are readily available to anyone to practice against, to analyze, and in many cases examine their source code. Chess engines are able to calculate far faster and many more positions than a human can, but they still lack the intuitive understanding or creativity an experienced chess player develops. They still have to rely on opening books hardcoded into them so they don't make idiotic opening moves and go into the midgame in terrible position. I suspect, in fair competition, AlphaStar has plenty of exploitable weaknesses, which the lead developers here admitted, and really we saw a few on display. They still withheld from or at least did not volunteer any information to the human opponents, and set up the matches to advantage AlphaStar in every way possible. Any exploitable weakness discovered in any one of these games, however, was immediately useless to the human opponent in the very next game, since the human opponent was up against a completely different AI.
I'm sure somebody will misunderstand me and complain, so yes, I fully understand that this was an exhibition match, and nobody at any time claimed this was set up to be fair. They did, however, try to imply it was fair as much as they could get away with, and this entire thing is and always was just a PR stunt for Google, the same as Deep Blue was a PR stunt for IBM. This exhibition match is an important milestone for artificial intelligence. There's some very interesting aspects to AlphaStar and what it had to learn how to do with a game like Starcraft II. I'm not going to try to take away from that, except to say that it's not a milestone of great significance to Starcraft II. Again, AIs will no doubt eventually beat even the best human opponents at Starcraft II or any other game humans design. It would be a mistake, however, to think that this has already happened. AlphaStar did not play the same game as the two humans did. It played an easier game, and it played only PvP which is frankly the least dynamic of all the match-ups in real Starcraft II (with its normal interface limitations), and on a single map. It also played against, with all due respect to TLO and MaNa, hardly elite players of the game, particularly with TLO off-racing. They are important names in the Starcraft II scene, and good choices for DeepMind to bring in early as training partners for their AI, but beating them is hardly something that will shock the Starcraft II scene. It will make a splash in the media where ignorant Wall Street investors will read about it, however, which was always the goal.
In a few years, and without a lot of fanfare, smaller, less well funded projects will start beating even the best human competitors in Starcraft II, and do so far more convincingly and without any tricks, just as chess engines eventually did some years after Deep Blue. Very shortly after that, they will be in the hands of Starcraft II pros to study and use as a training tool, which will dramatically improve human play, and Starcraft II AI vs AI matches will take up a small niche in the scene where a few hobbyists and enthusiasts will gather to watch. It will be all very ordinary, and it will be at that point that the real milestone will have been reached.
+1
Sam, da imas prevec tepcev, ki ne razumejo zakaj je level playing filed obvezen, ce zelis postavljati trditve tipa: AI boljsi od cloveka.
ZE dolgo nisem igral tega tipa iger, sam se spomnim, da je bil pregled in micromanagment kljucnega pomena. Da ne omenjamo bliskovitega odziva in izrabe specialnih sposobnosti enot, katerih pravocasna uporaba je lahko zlahka zagotovila zmago, ze sekundo pozneje pa je bilo prepozno in si ze izgubil.
Neumnost. Če je zmagal, je igral inteligentno. Postavi mu drug cilj in omejitve, zmago brez da igra tako, in bo našel drug način. Temu se reče inteligenca.
Predpostavljam, da ne igras SC2? Ker po tvojem je kalkulator pametnejsi od cloveka, ker zna 6 mestne cifre mnoziti med seboj hitreje kot clovek. V glavnem, ne strinjam se s teboj in mislim, da si storil fundamentalno logicno napako, najbrz zaradi pomanjkanja informacij.
Poanta vsega je, da se AI da priuciti, da je na nivoju (pro) playerjev. Kar je bilo do sedaj prakticno nemogoce, ne glede na igro. Brute force tudi vedno ne pomaga, ze sedaj je lahko AI goljufal in bil boljsi. Potem imas pa situacije, kjer AI postavis v situacijo, iz katere ne zna in se zacikla. Zdaj pa se zna prilagajat, ker ni hard codano.
Ceprav, prav v tej igri, kjer je MaNa zmagal, je bila taka banalna zadeva, ki bi jo AI moral prepoznat, sploh ce je imel za 200 let igralnih izkusenj (ali miljone iger za sabo), ne pa da izpade kot rookie mistake.
Ampak glede na tempo s katerim gre zadeva naprej, sploh ne dvomim, da bodo lahko ze cez par mesecev to lahko speljali na nivo, kjer tud pro igralci ne bodo kos vec nekemu uber stvoru, ki se bo znal prilagodit ne glede na situacijo. Nihce pa ne uposteva niti meta igre naspotnika (ker AI nima zaenkrat tipicnih slabosti), ki je tukaj ne mores nastudirat niti dnevne (ali sezonske) forme posameznih igralcev.
Gre pod crto samo za preboj magicne meje, kaj se lahko danes ze pokaze v praksi in ne toliko za to, da je AI kar sam po sebi pametnejsi ali boljsi.
Neumnost. Če je zmagal, je igral inteligentno. Postavi mu drug cilj in omejitve, zmago brez da igra tako, in bo našel drug način. Temu se reče inteligenca.
Predpostavljam, da ne igras SC2? Ker po tvojem je kalkulator pametnejsi od cloveka, ker zna 6 mestne cifre mnoziti med seboj hitreje kot clovek. V glavnem, ne strinjam se s teboj in mislim, da si storil fundamentalno logicno napako, najbrz zaradi pomanjkanja informacij.
Ni pametnejši, samo hitreje množi, bolje igra šah, go, sf2 in še marsikaj. V tej igri, če je neumna, pač ne more pokazat inteligence, v igri go jo je, igral jo je na način, ki je bil profičem čuden, in jih premagal.
Lahko rečemo, da DeepMind še ni boljši od človeških nasprotnikov v Starcraftu, ampak bo pa najbrž kmalu. Kar presenetljivo.
Nekaj kar zelo vpliva na Starcraft (za razliko od iger kot šah ali go) je poznavanje nasprotnika. Ko bo AI znal igrati na vseh mapah in z vsemi rasami in ga bodo lahko nasprotniki malo naštudirali, bo seveda drugače. Plus, njegova nasprotnika nista bila ravno top tier, čeprav sta zelo dobra. Ampak to bo čez nekaj mesecev, ko bo DeepMind tudi bolj pameten in znal več.
Sploh je zanimivo, da je bila glavna prednost AI odličen micro in ne toliko strategija. Bot, ki je imel opcijo pregleda nad celo mapo (in zmagal 10-0) se je je uporabil mass Stalker. Kar je načeloma manj učinkovito, kot uravnotežena kompozicija vojske. Ampak odličen micro (zaradi pogleda nad celo mapo naenkrat) mu je omogočal, da je napadel iz treh strani naenkrat in zelo učinkovito ciljal nasprotnikove enote. Tega človek (zaradi fizičnih omejitev) preprosto ne more početi.
Rekel bi, da bo čez eno leto DeepMind zmagal svojo prvo profesionalno tekmovanje
As we saw in that last game against MaNa, AlphaStar wins not by being smarter than a human, but by having greater capability than a human in terms of what it can see at any given time, and its rapid precision control. Just mechanically a human needs to be looking where he needs to be looking, then move a mouse to acquire a unit or units, press a button, move the mouse, and click. Humans beat each other in Starcraft II by having better situational awareness, and utilizing guile to exploit their opponents lack of situational awareness. I would be a lot more impressed with AlphaStar if it was actually playing the same game as its human opponent. That means actually looking at only one part of the map at a time, OR glancing at the minimap (humans can't do both at the same time). I know too much about AI and computer science to be fooled by DeepMind's excuses about focus. AlphaStar is still effectively cheating at the game, moving its "focus" to where it already has complete precise information, and then acting on that information based on decisions it's already made about that perfect knowledge in that area of the map
To je napol res. AI še vedno ni videl skozi fog of war, ampak je samo imel pogled nad celotno mapo istočasno, torej ni omejen na samo en ekran. Ampak izboljšan AI bo čez par mesecev samo preprosto premikal ekran dovolj hitro, da bo podobno zmogljiv, kot je bil trenutni bot, ki je "goljufal".
"božja zapoved pravi; <Spoštuj očeta in mater>,
ne govori pa o spoštovanju sodstva."
Janez Janša, 29.04.2014
Ko bo AI znal igrati na vseh mapah in z vsemi rasami in ga bodo lahko nasprotniki malo naštudirali, bo seveda drugače. Plus, njegova nasprotnika nista bila ravno top tier, čeprav sta zelo dobra. Ampak to bo čez nekaj mesecev, ko bo DeepMind tudi bolj pameten in znal več.
Ko bo AI znanje iz Starcrafta znal uporabiti v drugih igrah, bo upravičil naziv.
Sploh je zanimivo, da je bila glavna prednost AI odličen micro in ne toliko strategija. Bot, ki je imel opcijo pregleda nad celo mapo (in zmagal 10-0) se je je uporabil mass Stalker. Kar je načeloma manj učinkovito, kot uravnotežena kompozicija vojske. Ampak odličen micro (zaradi pogleda nad celo mapo naenkrat) mu je omogočal, da je napadel iz treh strani naenkrat in zelo učinkovito ciljal nasprotnikove enote. Tega človek (zaradi fizičnih omejitev) preprosto ne more početi.
A ne bi bilo bolj zanimivo, če bi bila glavna prednost ravno strategija in ne micromanagement?
