Only possibly because its an interesting topic, now you just need at least 1 guy who has a opposite opinion and knows how to reason without falling back to flames and name calling.

I admit calling it cheating might have been a bit harsh in this context (AI vs AI). Anyhow, it was an interesting debate but i gotta get my coffee and do some work (yeah, freelancer and all that ,p)

Also 18 post topic now

Tis true Anyhow, i have linked an interesting article in my previous post so you can see where i come from with my views on this.

The question is, would the overmind still win against the same AI with a locked APM ? What about 2 v 2 where 1 player in each team is a human? (would be way more interesting to see AI's able to do that and you not knowing who is human and who not.....

Sepharo wrote on Jan 19, 2011, 20:55:
"Just shootin' the shit about an interesting topic" is a hard style to convey in text.

A 15-post thread without the usual 'PC gaming is dying' or 'evil consoles are evil' stuff is a rare treat these days indeed.

If your point is that putting interface limitations on AI developers (say, the aforementioned APM ceiling) to force them to concentrate on 'real smarts' like more adaptive behavior, that's an interesting thought.

Do note though that in the article they mention how a competing AI did in fact find a flaw in their algorithm and yet the Overmind was capable of devising a counter and winning.

Sho wrote on Jan 19, 2011, 20:50:
Note that I'm not argueing against you here - I agree with your posts -, I was just reacting to eRe4s3r's cheating assertion.

I'm not arguing with anyone. I know it always seems like that on a forum .

"Just shootin' the shit about an interesting topic" is a hard style to convey in text.

No no, you missed my point The reason i am not convinced of this method to develop AI's is that they are not playing with a simulated human-like interface behavior/restriction. The end result is thus an AI that wins against other AI's but once it has to play "at the same level" as a human it becomes quite probably a sitting duck. Because own pathing method means they micro each move order for each unit.

This makes it a cheating AI. Like when you play UT2k3 on Godlike. (which removes the rotation dampening for the views that bots have simulated to restrict their speeds and other things).

Or basically, i don't downplay their work, but innovative it isn't. It doesn't advance game AI's a single inch. Because coding an AI that wins against humans is NOT HARD. Whats hard is coding an challenging AI that is unpredictable, makes errors, and can be won against by exploiting said errors. (Racing games are trying to get this right for nearly 20 years now). If you want to read on this, i would advise you to read the emergent AI series - part 1 below.

Link -> http://christophermpark.blogspot.com/2009/06/designing-emergent-ai-part-1.html

Note: Because of my affiliation with said game/developer my view may be jaded

This comment was edited on Jan 19, 2011, 21:00.

Note that I'm not argueing against you here - I agree with your posts -, I was just reacting to eRe4s3r's cheating assertion. For a fair human-vs-AI competition you *would* arguably decree an APM ceiling (even though, as you pointed out, the game likely cannot be won by mere APM brute forcing, so as a cheat it wouldn't necessarily get you all that far), but holding one never was the objective here.

The API can set the APM. I didn't see any mention of this being done for the tournament though.

There are quite a few rules dealing with the performance of SC while being interacted with. I'm sure high (for a bot) APM causes issues or just isn't possible.

I would also assume that a minimal amount of actions is preferable to wasted or unnecessary actions. I guess what I'm trying to say is that just a high APM alone does not secure a win.

I was thinking about APM, yup.

eRe4s3r wrote on Jan 19, 2011, 20:10:
Hence brute force, the Ai wins because it can do stuff at a faster rate than humans (microing) - it has more computing power, and can do many things at once at different locations.

Not really a fair competition when humans are involved Of course ANY AI would win if it micros units "intelligent" at 100% cpu power with a 6 core 3ghz cpu. This is the inherent problem with making real time AI's that can play "fair" against humans. And why most of them are "limited" via some kind of max APM and max "attention hotspots"

Obviously AI vs AI is another matter (so in this content it isn't cheating per sé, but humans playing against this AI have no chance to win once the dynamic strategy responses are of a broad enough scope.)

I think you're missing the whole point. It's about strategy. "ANY AI" is not going to win. The game speed is limited by build times, move times, and resources. The only advantage here is the speed at which commands are decided upon and entered. And since those commands are limited by the previously mentioned gameplay mechanics the only microing that's going to be done is the move orders.

Do you think if they simply applied their pathing to the default SC AI it would become unbeatable by humans?

Just seems like you're needlessly downplaying their hard and innovative work.

That was a great article on the Berkeley Overmind. They used some straightforward but robust methods for determining which units to attack, and made great use of the pc's ability to micro far beyond even the top level human players. I can just imagine mutas that could scatter to avoid area affect damage and were constantly dancing in and out of attack range would be absolute hell to beat.

Sho wrote on Jan 19, 2011, 20:05:
Note that this was made for a AI-vs-AI tourney, not to play against human players, so there's no cheating in going beyond human interface limitations.

But they did play it against human players because the in-game AI wasn't a challenge in the least.

If the human interfaces you're referring to are brain+hands or keyboard+mouse, then yes they obviously go beyond them. But the way it's interacting with the client is the same.

true enough, but i was more thinking of the "let humans play against it online"

Hence brute force, the Ai wins because it can do stuff at a faster rate than humans (microing) - it has more computing power, and can do many things at once at different locations.

Not really a fair competition when humans are involved Of course ANY AI would win if it micros units "intelligent" at 100% cpu power with a 6 core 3ghz cpu. This is the inherent problem with making real time AI's that can play "fair" against humans. And why most of them are "limited" via some kind of max APM and max "attention hotspots"

Obviously AI vs AI is another matter (so in this content it isn't cheating per sé, but humans playing against this AI have no chance to win once the dynamic strategy responses are of a broad enough scope.)

Note that this was made for a AI-vs-AI tourney, not to play against human players, so there's no cheating in going beyond human interface limitations.

eRe4s3r wrote on Jan 19, 2011, 19:49:
Interesting indeed - but still far away from a proper agent based emergent AI that can play the whole ability range of the game competitively.

Also interesting because they mention that Starcrafts path-finding is so hilariously broken that they had to write their OWN pathing logic from scratch. Which imo is cheating. The AI does not play the same game humans play, so it always wins when there are no exploitable loopholes left.

In a sense, this is also the "brute force" approach

They're still only interacting with the game through the same calls a human would/could be making. Of course at much faster rate, but the point of the agent is to play the game according to the rules.

It's their own pathing logic micro'ing move orders. I'm not a SC player, let alone expert, but I'm pretty sure pro players don't just click a unit and click across the map and let the pathing do it's thang.

http://eis.ucsc.edu/StarCraftAICompetition
http://eis.ucsc.edu/StarCraftRules

This comment was edited on Jan 19, 2011, 20:00.

Interesting indeed - but still far away from a proper agent based emergent AI that can play the whole ability range of the game competitively.

Also interesting because they mention that Starcrafts path-finding is so hilariously broken that they had to write their OWN pathing logic from scratch. Which imo is cheating. The AI does not play the same game humans play, so it always wins when there are no exploitable loopholes left.

In a sense, this is also the "brute force" approach

That Swarm article is really cool. Well worth reading.

Creston