Slashdot Mirror
Most Impressive Game AI?
togelius asks: "I have the feeling that when developers make the effort to put really sophisticated AI into a game, gamers frequently just don't notice (see e.g. Forza). Conversely, games that are lauded for their fantastic AI are sometimes based on very simple algorithms (e.g. Halo 1). For someone who wants to apply AI to games, it is very interesting to know what AI is really appreciated. What is the most impressive game AI you have come across? Have you ever encountered a situation where it really felt like the computer-controlled opponents were really thinking?"
Personally, while enemy AI is something that's pretty neat to see in action, it's the friendly AI that gets my attention. Most games seem to put all their effort into the enemy, while you friends turn out to be schizophrenics with an IQ of about 40. I haven't played, but I have heard that something that people complained about in Gears of War was was the poor team AI.
I don't play many games any more, but Halo 2 was one that I thought pulled ahead of the pack a bit. Friends that can drive vehicles was pretty cool (albeit not always the safest drivers...) allowing you to man the gun in the back. They also seem better at not running right in front of you when you're in the middle of launching a rocket, and also do little things like take advantage of available cover (or in other cases jumping up on top of said cover and getting blown to bits). Halo 3 is supposed to have even better AI for both friendlies and enemies, and that's one of the things about it I'm looking forward to.
"What do you despise? By this are you truly known." --Princess Irulan, Manual of Muad'Dib
/)
GalCiv II has given me the worst beatings I've ever taken in a strategy game. Shogun: Total War managed to spring a tactical ambush on me once (although in fairness my grip on tactics was much worse back then). Both of those AIs gave me a challenging game experience as a player, which is what "good" AI should be judged by.
However, if we're talking about "impressive" AI, nothing I've seen in the gaming world can compete with Paradox's EUIII. Yeah, I know, each individual AI nation makes a lot of bonheaded moves. But the game is managing the armies, navies, economic, religious, colonial and foreign policies of up to 300 nations, every game day when a game year can go by in a minute or two, on a 1.9GHz processor. Considering the number of cycles and the amount of memory avaiable for each AI opponent, it's simply amazing to me. I really think that should be the basis of comparison, not so much the level of play the AI achieves, but the level of play it achieves with the resources available to each AI player. If nothing else, that standard makes it meaningful to compare old games against new ones.
The important think to know about AI in games is it's not "AI". It's scripts or code that simulates scripts. There's no neural nets or anything else because we can't get the power for a neural net in an active game. In chess we can but then chess no longer is fun unless we tone down the "intellegence".
Hmmm. If I write a neural net program, how is that different than what you call "scripts or code"? It's still just code.
And the best chess algorithms (which you seem to claim are "real AI") are just search algorithms that search 30 moves into the future (with pruning)
You're making a distinction between "real AI" and "not real AI", when really there is no distinction. If a system can solve a problem intelligently then it's AI, regardless of the algorithm.
There *is* a distinction between "human-like AI" (neural nets) and other kinds of AI, but we'd be foolish to assume that human-like intelligence is the only kind of intelligence.
That's why the "best" game AI isn't necessarily the smartest or most responsive - it's the most human.
Writing an AI that makes the occasional "human error", or responds in a reasonable time is harder than writing the "best AI possible", but makes for a more believable (and of course, enjoyable (since who likes getting beaten all the time)) game experience.
To be fair, the fact that patterns work simply means the AI is deterministic. Pacman could have extremely complicated AI but if there's no learning from a players past, random elements etc you're always going to be able to learn to find a patten which works.
To be honest, I've always thought that AI in computer games sucked. Games are usually made hard by having the bad guys have better fire power, shields, energy etc than you, or having loads of them against one player. It would have been a laugh, for example, to have a doom style game with one player against one computer bad guy, but have him be as smart as a human. Thankfully, online multiplayer games mean you are no longer restricted to whatever crap AI system the programmers manage to string together, although the problem has now shifted to dealing with people cheat - a problem which games programmers show no signs of being any less inept at dealing with than with AI.
If we are talking about a computer game I'm hoping we are talking about human-like intelligence.
What I really was trying to get across was a common misconception (one that stuck me when I got into a game company) that "AI" as it's taught in school is very different hen AI as it applies to most games. The biggest difference is most AIs don't learn, and most are pretty much just a script that doesn't change. We don't have the ability to throw away any cycles of the game so the AI tends to be highly stripped down to the point it's just "oh I see a gun, I'm going to react to the gun, how should I react to the gun, I'll do that." This is completely scripted to the point where you can tell what's going to happen if you point the gun at the person a second time or a third time. There's no "thought" or "intelligence" to the system, thought it might seem "intelligent"
A chess AI on the other hand evaluates all the options of what it can do and chooses a best option, the pruning is a form of "thought". A chess master will be doing something similar where he thinks of all his possible moves and then considers responses and so on which is effectively using game theory. To me that's actual intelligence even if it's not fancy.
The difference between a neural net program is it's code that tries to simulate the learning and thought process if you will, the code that AI in games use is just like I illustrated above. There's an "action" and the code quickly decides what's the reaction and does it. It doesn't try to evaluate too much because we don't have the cycles to do that.
Unreal Tournament.
It had the first bots that you could play against for hours and not even notice you were offline. I havn't encountered a more convincingly human AI in the dozen FPS games I've played since, including UT 2k3(which probably means that the UT maps were just easier to code for).
It's the only game where you can feel yourself increasing in skill over the course of a few days of playtime, and ratchet up the difficulty a bit and get the same kill ratio, without feeling suddenly overwhelmed by perfectly aimed headshots.
People in Soviet Russia, however, appear to be afflicted with amusing juxtapositions of the aforementioned situation
Far Cry has great AI. I don't remember ever going "zomfg the computar is thinking!!!", but the enemies respond in a sensible, coordinated manner -- sweeping areas where disturbances are heard, covering each other, and using weapons intelligently. Far superior to Half-Life (and, credit where credit is due, several years after Half-Life), but also in a much more open-ended game world, where the mobs don't just have to navigate a room full of crates.
The really sad part was when, halfway through the game, they switched from "smart" soldiers, to Doom-esque mutants who just ran straight at you. Ruined that game for me.
Game... blouses.