{The List} AI

[Fosse]

I hate when I crosspost and the other guy's is way better.

[Nikolai]

A crosspost 13 hours later?

[wrylachlan]

Originally posted by Jamski


This ruined the MoO3 AI. It tryed to spread its forces over many fronts, meaning it was never a real threat on any of them. The Ai should think TARGETS, not fronts. It should have a goal like : get Athens, and defend Rome with x troops. Simple.

-Jam

That's not a problem with an AI thinking in terms of fronts, that's a problem with an AI being bad at thinking in fronts.

[Kuciwalker]

However, targets are how many humans think - so it must be good for the AI.

And I like how the AI was actually different on different levels in GalCiv.

[POTUS]

Originally posted by Nikolai
POTUS: If you can beat Deity, you're NOT the worst Civ player ever, I can assure you!

I've only done it once, unfortunately. All the AI's were peaceful and had no military. After that I discovered that I could select civs that were good, and not just have it random.

[wrylachlan]

Originally posted by skywalker
However, targets are how many humans think - so it must be good for the AI.

And I like how the AI was actually different on different levels in GalCiv.

Maybe I'm inhuman, but I think in terms of fronts. The Greek front, the Russian Front, etc. Within Fronts, I think in terms of targets, like "I want to take sparta" or "it looks like Catherine is getting close to Dover, gotta stop her." But my overall view is, "How am I doing on the Greek front? How am I doing on the Russian Front? How should I apportion the distribution of new troops?"

I guess I'm thinking like this.
1)Analyze each front.
a)How am I doing I doing at achieving my goals for this front?
b)How many more troops would I need for each of the goals?
c)Can I shift troops around on this front to make it happen?
d)Does this front need reinforcements to achieve its goals?
2)How do my fronts compare?
3)Reinforce the neediest fronts with new troops.
4)Reassess and possibly move troops from one front to another.

The reason I think this might help the AI is that it narrows the event window so that the AI doesn't have to figure the "what if" of a specific unit crossing the map to aid in winning a specific target.

All it needs to think about on each of the fronts are the targets, its chance to get them, and how reinforcements would change that percentage. Each front deals with reinforcements exactly the same whether they're new recruits or sent from another front.

Put another way it would be like having multiple AI's each Front is its own AI which executes its targets and requests troops or offers excess troops to the CIV AI which apportions new troops.

I think the modular approach is the way to go.

[MrBaggins]

"Fronts" are another simple to think about, tough to actually model, concept. Humans have an intuitive sense of grouping in a 2d space, and can ignore "clutter" relatively well... so a human can distinguish relative patterns, or troop fronts quite easily.

A machine can't, instinctively. The problem is actually fuzzy and not crisp. To solve the problem, typically you use an influence map. An influence map works quite well for single entities per tile models.

The complexity increases when you wish to model multiple entities per tile, of varying power. You can use scoring and essentially sum the values, to associate them.

Then comes the last challenge... prediction. To deal with a threat, you need to move units before that threat arrives... (well... not with unlimited travel railroads, but this isn't a guaranteed situation.) You thus have to predict where a threat is spacially, and how its shaping up. To do so you must propogate the value of each threat, from its initial position, in all possible directions, and decay the value, over a number of iterations of its movement distance.

This actually gets fairly costly in computational terms. It scales badly with map size and number of entities.

[Shogun Gunner]

Good to have folks knowledgeable about programming reading these threads. Many things seem possible until actually try to do it!

I've seen that responsive from the DBA and programmers in real life....its usually a polite "no" with a bit of scarcastic "do you have any idea what that will cost/size of the hardware system/lines of code necessary to do that?"

Glad you are here to keep us in check MrBaggins.

[Fosse]

Originally posted by Nikolai
A crosspost 13 hours later?

err....

[Shogun Gunner]

Isn't there a statue of limitation on crossposting? Has to be within x amount of time?

[Fosse]

If it makes Shogun Gunner look foolish: within 15 seconds.
If it makse Fosse look foolish: 10 days.

[Shogun Gunner]

ouch!

We've all x-posted...and also been a victim of x-posting.

Really something should be done about this issue!!!!

[okblacke]

I'm going to make the assumption that Civ 4 won't be giving us revolutionary AI--nothing heuristic, adaptive, neural netty, etc. I haven't seen anything to indicate those things aren't prohibitively expensive. (I haven't seen anything to idnicate that they ever won't be prohibitively expensive, but as long as Moore's Law holds, there's hope.)

Having said that, what would be incredibly cool is for the AI to be an external program. This would tie in with some other folks' request for the next Civ to be based in MP. (For myself, I wouldn't want anything to compromise the single player game, but this may help achieve both.) It could also simplify the whole "multi-threading" issue.

What I'm thinking is, put someone on AI with the problem of: Here's the data you'll get from the program with every change, now solve the problem with an eye toward various goals.

I don't want to go overboard with detail, but an advantage of treating the AI like a remotely connected player is that user mods could be developed. For most games that wouldn't be a possibility, but Civ has enough die-hards that it's almost a cinch.

It would even be possible for people to try out various AI techniques with Civ as a base. Make for more than a few great graduate theses.

[ok]

[MrBaggins]

Heuristics essentially means evaluating and reacting to feedback. E.G. Defending when attacked, stoping production of settlers when there's nowhere to settle, etc.

Any AI worthy of being called an AI does this.

[Kuciwalker]

If it doesn't use a heuristic, and doesn't use a neural net, wouldn't that mean either a) it moved randomly or b) it created a tree of all possible moves and evaluated them?