A History of AI Research in StarCraft | AI and Games #26

You should play Meepo for DOTA 2

Man...I really want to play Total Annihilation now..

Fascinating! StarCraft is such a good game. Keep up the good content mate

As a programmer whose computer science education was 30 years ago (when it was first being called that), and gamer, I really enjoyed your presentation, even though I don't do AI research, and don't play RTS games.

Nice video! Very interesting, subscribed.

Er, in the 2017 Human vs Bot match, the Human player actually lost to 1:3. He won the first one, then lost the remaining matches.

It's an amazing video. For once YT recommended something I want to see.

nice videos

Anyone else taken aback by the sudden switch from a Scottish accent to a really good French one?

Best non-korean protoss player. I laughed everytime i hear someone being described like that

Stork (Song Byung Goo) does have trouble with a certain advanced AI though. It's called Flash.

When your AI can play Dwarf Fortress and halt death of your fortress for at least five in-game hours I'm impressed. Other than that, get out! Plebs.

That Ben Weber guy. Ugh is he fugly.

Completely off topic, but what font were you using for the titles in this video?
It looks like Neutra, but the uppercase E's in your font look way better.
Nice choice!

As good as SC was in game mechanics as bad was in pathfinding. The real milestone it set was sightline rules. I don't think the AI used it really, tho.

yeah because protoss is ez thats why they use protoss

I was amused to know that the best BW AIs do not come near close the korean pro player, I thought the AI had a big advantage over the human because you can select so many units at a time, a limitation an AI dont have. I guess this just show how cool our brains are, and how difficult really is programing all this. Excellent video.

How fast can the Deepmind play sc2? Maybe the difference between it being able to beat other games vs Starcraft has to do with the speed at which it can simulate the games. In the Dota2 video they said theirs had "lifetimes" of practice over a number of days - but it doesn't sound like that is the case with Starcraft? Thanks for the video!

Man, I feel you about the DOTA 2 comment!

pretty informative video. thanks

I wish I'd thought of this when the video came out but I guess now I cross my fingers and hope you notice it, because I think this is a pretty interesting concept. So sc2 Archon mode is quite fun with two players having to work in conjunction, one taking over micro, the other macro. What if you could do that in any game, and the other player was an AI. What if we had a competition similar to the style in the video, but utilising a combination of AI and human input. Humans can be either on offense or eco, and the ai has to be a complementary variant. It cannot be the main driving force. This is more to see whether we can make AI that complement the player in interesting ways, like for example Age of Empires 2 with ai controlled monks, or sc1 with automated scouting.

There is a way better game to train and develop ai then StarCraft

As an Alberta boy I cannot help but be proud that people In my own backyard pushed studying RTS AI and making SC bots haha so awesome

My life for Aiur!

The Alphastar games last week or so were pretty interesting. Looks like they have a fair amount of work ahead of them, even though they won (10/1). I was hoping you would have mentioned the conundrum of how Blizzard themselves handle the SC AI's.

Well, for the record, in january 2019 AlphaStar defeated Mana in TvT, so, did'nt take too much time after all.

Hello, im from future, pls stop Alphastar while you can !

I played starcraft 1 way too much. At least 100k games. I would love play against one of those Starcraft 1 bots.

I started to develop an AI bot, but after installing Java it kept telling me to update, so I gave up.

Aiur is the name of the protoss homeworld and its pronounced I-UR there i said it

Just found your channel, loving your work, keep it up.

When did, if ever, Firebats cause splash, collateral, damage in the future of StarCraft games?

Funny story, over a decade ago I was playing SC1 against a hard protoss AI. The match went to late game and for no reason the AI thought a viable strategy to end the stalemate was to fill the map with hallucinated archons

You can’t make an AI play Zerg and not call it overmind hahaha

I think rts is a really interesting area of AI research since it's imperfect inmormation and practically unlimited gamestates. But the game computers and humans play are fundamentally different for a human the biggest barrier to improvement is speed. For humans starcraft is not a strategy game it's a multitasking simulator ie how many things can you do at ones. As someone who's climbed the leagues in sc2 the biggest difference is just how fast im playing. I can beat weaker players with next to no strategy since they simply can't keep up with me in speed

This is fundamentally different from how AI approach the game. For a computer starcraft is almost a turn based game. A computer can execute strategies that a human could never do, because of physical limitations. I can't click that fast and accurately no matter how much I try. This is what AI researchers fail to understand.

You can artificially handicap the computer to level the playing field. But at the end of the day it's never gonna be the same limitations a human faces. So in conclusion AI and humans are playing two separate games.

"How hard can it really be?"
Every player worth mentioning is a PHD
"Oh."

5 years later

Hope all the StarCraft fans out there enjoy this new video. Ever made a bot? Fancy trying it out yourself? Let's hear from you.
And yes, I mispronounce Aiur. I have since learned my lesson. 😀

CORRECTIONS: The BWAPI was originally developed by Michal "Kovarex" Kovarik and later handled by Adam Heinermann. Apologies for any confusion.