PDA

View Full Version : Any DECENT Strategy?



Antoninus
November 21st, 2008, 03:38 PM
One constant failing I've noticed from the RTS arena is the AI.

Often cranking up the difficulty on an RTS game just means the computer cheats more than usual and there's no real STRATEGY involved.

The computer just builds masses of crap and hurls it at you like an eight year old in a snowball fight. Every now and then you get a game that utilizes a few special units in a strategically effective manner but often its more random chance that makes these units useful. Defeat usually comes because you didnt build enough turrets.

A strategy I've found to be almost fool-proof against AI players is to simply build a good enough defense to weather the constant barrage of frontal assaults then concoct some sort of sneaky flanking maneuver.

A CPU has limitless patience and I've sat there for hours while a CPU player hurled assault after assault against my defenses and each and every single time it was repulsed with heavy losses. The CPU NEVER tried anything different. NEVER changed it's strategy, NEVER tried to build a bigger attack force, NEVER tried any other strategic moves. Just a continual all-out frontal assault.

Why is it so DIFFICULT to have a CPU try something, see why something failed, then NOT try that again until a future date?

Now take this with mountainous piles of salt as I am not a programmer nor do I have much of an idea of what goes into programming an AI player.

Are there any RTS games that use strategy to any real effect?

At this point I'd settle for Total Annihilation graphics if it meant the game had AI that wasnt pants-on-head retarded.

staticonthewire
November 21st, 2008, 05:22 PM
Simulation AI is difficult to do under the best of circumstances. If on top of that you're constrained to what you can package into a game, the odds of getting something really good are slimmed down even further. Most games go for prepackaged scenarios rather than procedural AI, because the game has to respond quickly to the player. Prepackaged AI does make it all a bit stale. But at least your game responds in a time frame most players would consider reasonable.

One of the most significant failings in much of what passes for procedural game Ai is the sterility of the evaluation functions the computer uses to determine success or failure. There is also a tendency to go for a "stateless" model, which is where you get the Ai running the same scenario over and over without seeming to realize that it might be worthwhile to cast around for a different approach.

It is an interesting topic though, and many people are working on improving it. A newly popular approach is to "prebake" a set of scenarios into the AI at a certain level of abstraction, and then include a way for the Ai to evaluate ad-hoc scenarios in real time and find an abstracted scenario that fits. This requires the game developers to play the game through, perhaps thousands of times, and record and analyze a lot of data. That sort of thing doesn't fit too many Gantt charts.

Remember also that the AI can't be TOO good. Game developers have to tread a thin line between ironclad AI, which in most games could be made simply unbeatable, and stale AI of the type you lament. This makes it even harder to design good game AI, since you can't really optimize it...

But there is hope. There are a few standard AI engines coalescing, and they will eventually make their way into your games.

But be careful what you wish for. You may catch yourself looking back in a few years at the simplicity of today's games rather wistfully.

Antoninus
November 21st, 2008, 05:29 PM
I'd love to see an ironclad AI

Most games seem to substitute a good AI for cheating

staticonthewire
November 22nd, 2008, 12:17 AM
You wouldn't enjoy ironclad AI in a game. It's fine for calculating actuarial odds or doing risk analysis for a bank or assembling insurance cones for some Cayman Islands shipping company, but it stinks when entertainment is the issue.

It's the backgammon opponent that NEVER makes a mistake, the chess program that's examining moves a dozen ply ahead of you, the blackjack AI that calculates every last statistical nicety to the last millionth of a percent; in a game, it's BORING...

What you want in a game is faulty AI, emotional AI, slipshod but occasionally brilliant AI. In short, you want something sort of like... a person.