In a mostly grateful comment on my review of his Raging Tiger and Falklands War 1982, Curt Pangracs writes that I underrate the AI in those games when I refer to it as “lackluster”.
A few days later, veteran wargame reviewer Jim Cobb writes an editorial on Combat Sim that accuses developers of not spending enough time or energy on the development of competent computer opponents. (Full disclosure: I owe Mr. Cobb a review of Blitzkrieg II that I *promise* to get around to very soon.)
The “lackluster AI” criticism I levy at the ATF games is, I admit, wargamer boilerplate. Hell, I suspect that the word “lackluster” is found more often in game magazines than in any other press form. As I explain in my reply to Mr. Pangracs, my more general point was that the AI seemed well able to handle the expected and obvious, but not the creative. Cobb’s complaints about AI in are along similar lines, only he wants the AI to not just respond to creativity, but be capable of (or programmed to) surprise.
All of this raises the obvious question of what to expect in wargame AI. What makes an opponent believable?
There is one sector of wargaming opinion that holds that, since most real wargamers seek out human opponents, energy spent on the AI is wasted to begin with. To me, this puts the cart before the horse. If wargame AI was, in general, better, there would be less need or desire to seek out humans.
I have a more basic question, provoked by Pangracs’ reply to my review. Do we know good AI when we see it? And should we believe what we are told by developers?
Really bad AI is easy enough to recognize. It was very common in early sports management sims, where opposing GMs would never challenge you for big free agents. Early wargames had computer opponents that had a good sense of the mathematic value of objectives, but poor sense of geography. The latest offering from Paradox, Diplomacy, has multiple AI opponents, none of whom are sharp enough to cut soft cheese.
When a game brags about its AI, it’s never a good sign. The chaotically stupid Superpower games were promoted on their realism and “learning” opponent. Make a game complicated enough and it may appear that the AI is learning (people can convince themselves of anything) but even if the AI was good, the games are far too random to test how good.
But AI that ranges from OK to good is hard to detect. Most difficult computer opponents are just given more advantages. They “cheat” in order to provide a challenge. This is not greater intelligence, of course, so a challenging game is not a sign of a good AI.
Wargames AI seems easy to program. There are limited goals defined by the scenario. There are limited resources available and rarely a need to produce more (most “strategic wargames” like Grigsby’s World at War are strategy games to me, not wargames). Include a combat resolution table or a sense of depreciating supply assets through a mathematical thingamajig and voila.
Apparently not so easy. Even in a wargame as simplistic as Rome: Total War‘s battles, the computer opponent can be easily convinced to prioritize its General’s uber-power over the same unit’s importance for the preservation of the army. Result: suicidal generals who are easily destroyed.
So what do we expect? An opponent that plays by the historical rules is fine, even though, as Cobb notes, any human opponent who did things purely historically would be beaten because you’re not dumb enough to act historically when you attack him. A computer opponent who had more than one programmed opening and the good sense to know when stall an advance would work.
As I still struggle with Noble level in Civilization IV (I love games, but fear I’m not very good at them), I am reminded of one of the most enjoyable wargames I’ve ever played. Sid Meier’s Gettysburg. I won’t deny that it is more fun in multiplayer. Me and one of my MP arch-nemeses have many war stories to tell about the times he took a strong position on a hill or when I forced marched reinforcements through the woods to hit his rear. All great times.
But the computer opponent was more than acceptable. It seemed to know how to regroup, when to withdraw its guns and where to withdraw them to, could scout, would extend its line, would refuse its flanks in trouble…Sure, with practice I could beat it pretty soundly. But there was a lot of practice.
Does this mean that AI is not all that hard? Probably not. Meier probably had some tricks up his sleeves, or, like many gamers, I have chosen to believe something that is not exactly true.
So maybe we don’t really need better AI. We just need to be fooled better.
“Wargames AI seems easy to program.”
It does? I would think the exact opposite.
I mean in theory more than in practice. Turn based, lots of math, set the objectives, basic rules (don’t outrun supply, watch those flanks, air superiority matters).
Compare that to even your run of the mill grand strategy game. Lots of different competing priorities, changing nature of opponents, diplomacy.
Wargames seem, on the surface, to be more straightforward than many other games. And there is a general idea that straightforward means simple.
You and I (and Jim Cobb, I hope) know better. Maybe the very directness of wargames makes the failings of the AI just easier to notice.
Actually, I don’t even think it seems simple in theory. You’re taking maybe 200 or 300 individual pieces (with different capabilities), placing them on a grid of maybe 10,000 spaces (not equivalent spaces, and spaces that have different values in different circumstances), and asking to find the best solution over dozens of turns. I think that sounds pretty tough before you even get to thinking about how to program it.
This is going to be a blog post for me because there is a lot of stuff here to comment on.
I am mostly talking in general impressions based on the nature of the game. Sure, when you start thinking about it the picture gets clearer. But you look at a game like, say, Great Battles of Hannibal, in which the AI has no idea how to protect the flanks of its phalanxes. Compare it to even a garden variety RTS that requires a computer opponent to juggle economics and a unit production queue and the surface impression is of calculations infinitely more complex than a wargame.
Given that I still bring up the old “Realism” article, I’m already looking forward to that post, Bruce.
I tend to agree with Troy; AI seems easy to program, until you actually try it. Making a good AI, in the sense of something that beats a human, is almost impossible, except in “twitch” games.
I tend to doubt that this is what people want, however. “Difficulty” is highly subjective; AI that I consider moronic is frequently cited by others as being extremely hard. this particularly goes for “complex” games where the very complexity of the game often succeeds in hiding how poor the AI is.
Frequently, what people want is an AI that can provide a nice challenge – but which they can beat! The key, to me, then – is that the AI should avoid doing anything obviously stupid. As long as the AI behaves sensibly, then it doesn’t matter that it has no answer to my brilliant maneuvers. It just mustn’t be stupid (e.g., marching its 200K man army into a pointless mountain province to starve).
AI stupidity “breaks” the immersion factor of the game (that buffoon is supposed to be Caesar?). It destroys the enjoyment as well – you want the player to feel that he won because he was good, not because the game AI was stupid. And that, IMO, is really the key to what makes a good AI. It should ensure that the player enjoys the game. Providing a strong challenge, avoiding cheating, and all the other things cited are just the means to that end.
P.S. Not really rocket science this; it’s been said by many others before me. But worthwhile stating again.
I agree, but I absolutely hate it when you know that to make it harder the AI has just moved across the game in a matter of seconds. This is more to do with FPS though. Also, I think the Rome Total War AI is pretty good except when they are attacking a city, where all you have to do is destroy their battering ram and then they just stay still. I think the campaign AI is very good. Also on Alpha Centurai the AI is really good especially at diplomacy
Wargames AI are not easy to program.
Even in the simplest wargame, the options available for the computer program in its search for the best move is big enough to make other hard optimization problems to look like a child’s game.
I don’t expect wargames AI to get any better until developers start looking at sound academic research on optimization. If this option is too hard, we can say goodbye to the dreams of good computer opponents.
John Tiller and col. in his latest “The First Blitzkrieg” used a data-mining approach. It is an exciting start point.