Harstem Versus Best StarCraft II AI

[@Harstem]

Probots Season 3 is being played out this week. You can check out their live stream on the following youtube links: Semi Finals will be this Thursday Dec 12th 7:30 PM EST : kzbin.infoCvK42BJY91Y Finals This Saturday Dec 14th 12:30 PM EST : kzbin.infoRZNrYwpLzj0

[@yzraeloa]

"If you beat me I'll unplug you" - Harstem making enemies with our next overlords

[@jeetkunndo]

Next imbalance complaint by bots: Human are imba

[@almondbungames584]

Those bot overseers are like those nasty flies in your room that you never catch.

[@drimmlunk172]

I am an sc2 bot builder. But my game knowledge is about gold level. My bot's best opening is a proxy BFF spamming tanks driving right into your main. So my bot is about gold level too.

[@Vers-AI]

Hey all! I'm running the Probots Tournament where these bots compete, I am in the comments to help answer questions :D

[@LadyMorrigan]

Why is sharkbot a zerg cabal conspiracy theorist :D

[@Vers-AI]

@@LadyMorrigan I mean, I don't want to call out Sharkbot or anything but.... all myths have a line of truth 👁

[@yixing09]

I haven't been following the bot scene, but didn't OpenAI have a really powerful bot a few years ago? Does the tournament format specify that the engines have to be hand-crafted vs trained on the ladder or something?

[@emissarygw2264]

​@@yixing09didn't the open ai one have special access to data that normal bots don't? Thought that got revealed later

[@Vers-AI]

@@yixing09 you might me thinking of DeepMind, OpenAI did Dota. But no our rules don't specify or restrict how people develop their bots. Many bots will have pre-determined behaviour but some will use elements of Machine Learning or Reinforced Learning

[@hung409]

Seems like Harstem needs to play some GTA to get his "run over some pedestrians" fix

[@Untrimednes]

The bots probably suffer from hedging their bets too much in the early game. The random spines and gateways delay their pace which otherwise would be better than the player thanks to the bonus mining but that doesn't work out against a human that isn't cheesing, but in 1000 games against other bots those probably pay out on average so they go for those. The burrow play I'd image is extremely effective against other bots which should have poor object permanence and simply forget units exist when they burrow outside of vision.

[@pwoblo]

I was dying at the "inefficient rain dance"

[@imacarrot6570]

Caninana is a very careful slow bot that sort of blooms into insanity. I think the other zerg bots like Eris are quite fast at expanding and you can find yourself behind quickly. They all seem to have trouble with engaging into a strong defensive position like tanks. The get kind of wishy washy and hover around the edges and just die to attrition usually.

[@endersword4247]

Thanks for doing this, great addition to the Finals

[@INS1cells]

Irs crazy even with all the APM in the world bots still are not even close to competitive with pros even off race. Just to show how much nuance goes into the game and why we keep playing and watching after all this time.

[@kevas777]

Its just because these bots made by amateurs not any big company like Google/OpenAI/Facebook. If they want - they would do it easy to take down any pro gamer. There are AI for chess/go that win any human just because big company do it, but people still playing vs humans and make world championships in chess/go so if there will be some SC2 bot that will win any pro-gamer, humans will still play in SC2.

[@Fimbu1vetr]

​@@kevas777Alphastar was developed by Google's Deepmind and it didn't beat pro players when it had human limitations (only had access to 1 screen at a time). Alphastar that could micro/macro on the whole map at once beat Serral iirc

[@boringstuff1542]

@@kevas777 Chess and Go have infinitely less decisions you can make to change the tide of battle.

[@comestible6270]

yeah i wonder if alpha star was still being worked on if it could beat serral

[@Fimbu1vetr]

@@kevas777 alpha star was made by a big company and it could not beat progamers when limited to performing actions on 1 screen. Without that limitation, it beat Serral

[@sandrorass890]

Shows that Terran is imba, if even Harstem can beat non-terran bots.

[@zergrush8709]

WHAT WAS THAT FIGHT 7:22? he wasn't even using his autoattack. killed ~7 marines and a tank with 13 stalkers and 3 zealots. I don't think the zealots ever connected

[@endersword4247]

I think the Zealots there were afraid to leave the shield battery range and so kept walking forward then retreating. Bots are notoriously not great at choosing targets well, like knowing to get on a tank

[@jonbecherer5103]

THESE ARE GREAT, THANK YOU

[@Eliteownage]

Bots can't see the widowmine lines which is why the units don't get picked up.

[@FlannelPrime]

Do you know why they can't see them? They can see projectiles like stalker pr marauder shots

[@samfrostinjapan]

@@FlannelPrime It might have something to do with the projectiles being created each time as opposed to theoretically the mine's line being a visibility toggle of an already present object. Purely guessing though.

[@kesatola3554]

@@samfrostinjapan Can they 100% not see it? I could see it happening that the bots can actually detect it, but choose to ignore it, because the widowmine controlling AI could keep switching targets non-stop and by doing so, messing up the opposing AI.

[@朕是神]

If they can see which unit a siege tank is aiming at, they can see which unit a widow mine is aiming at, line or no line.

[@Eliteownage]

@@FlannelPrime These bots cannot see projectiles. They don't work the same way the bots in the game work, they are separate programs from the game itself and connect via an API that does not give information about projectiles and some other information like the widow mine lines.

[@martinprince8253]

The Harstembot is the deadliest of them all

[@shanewang3813]

I watched the round of 8 and Caninana was speed mining, weird that it did not do it in your match

[@Vers-AI]

When bots play each other they play at really fast speeds, but when they play against people they need to be reconfigured for "Real-Time" mode. This is the first time Caninana has been available for people to play so, what we're seeing is the first attempt at getting it to play Real-Time and likely their code for speed mining broke.

[@drimmlunk172]

The speedmining is turned off in versus-human for fear of overstepping the real-time bound, which would lead to chaos in the bot.

[@true_xander]

15:35 sudden Star Wars reference

[@pawe8765]

awesome. what happend with alphastar?

[@niquil7800]

Thanks Harstem Altijd fijn om weer een video te kijken

[@GY-bd9bo]

progamer vs programmer

[@bentosan]

Prograndma

[@Jcarroz]

Anyone else think Harstem looks like Aniken Skywalker here?

[@miza6]

He does make a reference to sand…

[@calebadam2576]

I read thia title as programmers st first and thought "no chance."

[@ledrash6079]

And here I was thinking that humans wasted their apm on dumb stuff that doesn't help them.

[@gobbelhuhn3830]

I was suprised we didnt see a raven from you in the 2nd game. Youre always encouraging its use in TvP, yet against a Bot relying heavily on burrow micro you didnt build it. The again the captain obv doenst need helpful tools against those ai fools :)

[@ErazerPT]

Nice vid, but in a way it highlights something akin to the way the best computer vision human engineered solutions got destroyed by relatively "simplistic" ML models. You get very precise "clockwork" that functions amazingly IF under the right (expected) conditions, but all goes to hell in a hand basket the moment it doesn't. ML models otoh, like Alphastar for example, will have built insanely huge "decision trees" and will (try) to react the best they can to what they have at hand instead of "sticking to their guns". Sometimes with admittedly wonky results.

[@WilliamKiely]

Why aren't semi-final bots better than pros? AlphaStar was created in 2019.

[@yjlom]

what about chimera players though? they dominate in chess but in starcraft with how different human and bot skillsets are it should be even stronger

[@aquilegia9734]

chimera players only dominate in correspondence chess, because there is effectively unlimited time to think/analyze. If the time control is anything close to classical chess (2h/40 moves, 1h/20 moves, +30min) or faster a bot will be heavily favored in the match against a chimera player. Starcraft decision making needs to happen in a fraction of a second. It would be very difficult to get any kind of synergy going between bot and human on such fast timescales.

[@b1rds_arent_real]

​@aquilegia9734yeah, but you can give humans speedmine, bunker gas, autobuild or a bunch of other enhanced QoL features that'd make them much stronger. Also, you can make a units lost tab given perfect memory. Autoburrow/autoblink could make some difference too. Build orders can be displayed in-game or performed by a machine too, so you only have to micro. I'd suspect a mid pro with automine beats a top pro off of sharp timings they can hit and humans can't.

[@aquilegia9734]

@@b1rds_arent_real So basically your argument boils down to: would Harstem with 12% more mineral income (speedmining) be able to beat Serral? Would be interesting to find out! We know maphack alone wasn't enough of an advantage for our captain... ;)

[@kiratt77]

IT annoys me that it wants to format C:\ it is not a drive but a path.

[@TripWithTheMoon]

All the bots I play, insta leave the game for whatever reason.

[@Caolboy]

The only thing I noticed was that Terran can beat Protoss and Zerg even if they have 20-50k APM. If that doesn't prove Terrans are imba, idk what does XD

[@atifarshad7624]

Pretty sure top players can beat these bots handily, even if thry play Zerg or Protoss

[@robezharvey]

Harstem, bro - You have to hyphenate "pro-gamer" I and everyone keeps reading this shit as "programmer" PLEASE CAPTAIN

[@mothrahlurker788]

skill issue

[@loudaccord]

Anyone else think that the bot switched to Zerg mid fight? lol

[@pauledwards2052]

Have not watched yet, but I assume Harstem completely outplays the bot in every way possible, and then loses anyway because 50k apm micro is op and shouldn't be allowed.