Hacker News new | comments | show | ask | jobs | submit login
StarCraft AI Competition Results Announced (ucsc.edu)
50 points by phreeza 2319 days ago | hide | past | web | 19 comments | favorite

More detail on the winning bot Overmind:


They had a team of 13 Berkeley AI researchers working on this bot.

Thread on Team Liquid forums (Starcraft community site as well as Starcraft team) about this: http://www.teamliquid.net/forum/viewmessage.php?topic_id=160...

Man, watching those mutas fly around really makes me wish I could personally interface with BWAPI. I've always wished it was possible to manage each unit individually, even in combat.

I guess that's what turn-by-turn games like Advance Wars are for though.

The winning bot's strategy seems a bit one-sided (mutalisk-heavy). I guess no bots were good enough to build real counters to that? Looking at the videos, everyone is trying to kill mutalisk with Goliaths, which is not exactly the strongest tactic.

Mutalisks are a natural choice. The AI can individually control units to a super-human degree. Mutalisks are a harassment unit with a lot of potential upside when controlled properly.

Mutalisks are cost/power balanced around a human's ability to use them. Since AIs extract disproportionate value from them, they're likely overpowered for AI play.

The effect of AI control actually turns some aspects of game balance upside down. For example, Archons are normally considered an effective counter to Mutalisks (because of their splash damage). But AIs can exploit the Mutalisk's slightly superior range:


In human play, Scourge are usually a decent way to handle large numbers of Mutalisks. Small numbers of Mutalisks can evade Scourge, but in large numbers they're forced to trade. Not when the AI gets involved:


Here are Wraiths versus Hydras (normally a cakewalk for the Hydras):


Goliaths are the standard response to mutalisks, as goliaths are Terran's strongest anti-air unit. They have a pretty good dps vs air, and they significantly outrange mutalisks.

The Berkeley bot's mutalisk micro was just too good.

Goliaths alone? I don't think so. Any player with good micro (and we're talking bots here) will keep dancing out the damaged mutas, since Goliaths can't follow to kill. You should at least throw a couple of Valks into the mix. A bot could also pull off the cloaked Wraith + medic combo, perhaps.

Scourge snipe valks since they cannot move between attacks #1 and 8.

Valks also cannot fire once there are too many sprites/actions going on. They may have been excluded because of that if you didnt see any.

Does anybody know how these bots perform against humans players? This seems to be early enough in its infancy that I would expect humans to outperform computer players pretty significantly but I might have too much faith in my own race.

I'm with the winning UC Berkeley Overmind team.

We played against Oriol (the guy from the video linked in another reply), many times. Maybe 20. We won once. Now, keep in mind, he is a retired pro, and so he represents one of the very best players.

That said, by the end, he definitely needed to be very careful. He couldn't win any way he wanted to; he had to attack at the right time, etc.

We'll be posting a video of him against us soon. Probably the one where we won. :-) http://overmind.cs.berkeley.edu/

Starcraft has quite a few more variables than chess and tight time constraints to evaluate position, plan, and act so I don't expect Starcraft bots to get as clever strategically as Deep Blue. On the other hand, bots should be capable of flawless micro.

Is that enough to win a series against a champion of a major tournament? Is there a silicon BoxeR among us?

Flawless micro is an enormous, essentially insurmountable advantage, if backed up by even a modestly capable "strategic" engine. Even the world's best starcraft 1 pros make many micro mistakes over the course of the game--I'm talking blunders so bad that even I notice them. This is because a human cannot be everywhere at once, while an AI can. Take all the "mistakes" away and what you have is something that is nearly unbeatable without doing some sort of cheesy all-in strategy that gets lucky.

I think the true test of a Starcraft AI would be one with human-like restrictions. E.g., only so many actions allowed per minute, only vision of the current screen, built-in delays for actions that would tie up a human to perform, etc. Then the contest is truly more about intelligence and less about brute force.

I'm not sure it is insurmountable. I would love to see Flash play against AI bots a la Kasparov.

Maybe your reasons are the ones why we will never see a starcraft AI superior to human ones. But I suspect the real reason is more closely related to the disparity in money and minds dedicated to the two tasks.

In a straight up game, if the human plays standard then the computer will have a pretty good chance.

The main issue is stracraft is pretty complex and if the computer is predictable, a human can just exploit that for easy victory. (always 6 pooling for example)

Yes. Although build orders are not a very good example --- because that's on area where computers can be made more flexible and random without too much difficulty.

I guess that bots' micromanagement will become superior to humans, if it isn't already. A bot won't have any problem coordinating three or four battles at the same time. For humans it's quite hard to fight a multi-front war given StarCraft's UI.

This is what I've found.

Even the Insane AI in SC2 (which gets cheat advantage to minerals and gas) can be beaten 1v3 with just a 10 population protoss.

AI vs humans is rarely the standard 'Starcraft game' so much as its a puzzle game. The winning thing vs 3 insane AIs in SC2 was merely to forge-rush. Once you knew how, you could never lose to that AI again.

I'm sure the same is true of these AIs.

From TFA:

>While the expert player was capable of defeating the top performing bots in the competition, the results are quite encouraging.

Are there any plans for the winner to assimilate the loosers, as it wasn't hands-down the best?

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact