http://arstechnica.com/gaming/news/2011/01...competition.ars

the more i read this the more i envisioned some kind of military assist system. Especially when they mentioned building a threat map and using that to route various units around.

But the attract/repel system for handling threat evaluation is also interesting, especially when they added a self-adjusting component (run 300+ rounds of the same enemy inside a corral and by the end if it the threat level of a type of enemy has been adjusted accordingly).

How can this go to a Starcraft bot?

Has anyone played against a Supreme Commander 2 Ai? Those things are ridiculous, and they keep getting better.

Hard problems that this doesn't solve: threat evaluation when you don't have stat tables for all combatants, computer vision, unreliable units (weapons can jam, soldiers can get distracted, etc.), or for that matter even in-game pathfinding (their technique is vulnerable to local maxima, but it just so happens that because flying units don't need to route around anything they're immune to the issue).

It's an awesome achievement, but looking at this as a "first gen tacnet basis" is like, I don't know, looking at hip replacements and pacemakers as "first gen cyberware basis"; in some sense it's correct, but not in any useful way.


What? How can it not, given that it's a Starcraft AI competition?

~J

I found the article interesting for other reasons.

Hm... know we know how Mirage will be born.

Even if you don't have stat tables, you can infer some information given the past results. Basically, you will be adding some sort of fuzzy values to a table and work from it.
Pathfinding is always a problem, but as I understood, they didn't delve too much in pathfinding because they had little time to improve the built-in pathfinding system.

Actually you'll find the opposite. The built in pathfinder is so terrible they never used it.

That's what I meant, they didn't have enough time to improve the pathfinding the game had (because it was a pile of crap and no one wanted to touch it).

I just didn't add the part inside parenthesis

But that kind of learning is an entirely different problem which isn't even attempted (and didn't need to be) by this bot.


What? They can't improve the game's pathfinding—they don't have the source code, they interface with the game through a hook someone was able to make using DLL injection. They wrote their own pathfinding, by which I mean that when the bot decides "this unit over here should go over there" it doesn't issue a direction to the game for the unit to do that (and thus let the in-game pathfinding figure out how to accomplish that), it issues a direction to the game for the unit to move a small distance along a route it makes itself and keeps doing that until the destination is reached.

~J

Nonsense on two levels. One Kagetenshi pointed out. Two: if you don't have time to improve a pathfinder you don't have time to write one.

All issues that would be worked on would it really be used as a tacnet system.

Threat evaluation can be done on the fly based on reported observations, lidar (better then some kind of camera system at this time, imo), maybe one can even fingerprint weapons based on noises made. Sure, one can not have stat tables for combatants, but their weapons can be statted (tho concealed explosives could be a issue).

The first gen or two may be more strategic then tactical, as they could (or perhaps even already exist) help evaluate a rapidly shifting combat environment with data feed in from spotter drones, sats and various other sensor systems.

But my point is that while this is legitimate research, what they've done is still much easier than all of those problems "that would be worked on". To give another shot at the illustrative analogy approach, it's like pointing to a successful rubber tire and declaring it a "first-gen car basis"—it's a step forward, but such a tiny step compared to the overall set of problems that it's really not a reasonable description.

~J