Why Great Strategy Games Need Great AI Generals


When I was a kid, my friend and I used to play Command & Conquer: Red Alert together almost every other night over a dial-up connection. Mostly, we played against AI opponents, trying to take on more and more of them as we got better at the game. The catch was that the AI was incredibly passive: Save for a few small raids, AI players would happily build ever-larger bases and armies to no apparent end. However, the moment one of your units attacked their harvester, the AI would attack relentlessly with everything it had. It was incredible to watch happen on the minimap: In the blink of an eye, the entire map would seem to lurch toward whichever player had committed the offense. It was basically the They Are Billions of the 1990s.

This is a companion discussion topic for the original entry at https://waypoint.vice.com/en_us/article/j5n89x/why-great-strategy-games-need-great-ai-generals


Age of Empires II has been the go-to strategy game for me and my circle of friends basically since it’s release. We basically only play comp stomp, so when they patched the AI for the HD version, that really messed us up. Previously we could take on a full team of enemies on hard difficulty without even breaking a sweat; now they would stomp us even on medium. It was, honestly, kind of fun. For the first time since the release of the original game (20 years ago?) we had to re-examine how the game worked.


I wonder if smart AI is always necessary in strategy games. Fire Emblem would be ridiculously hard if the enemy AI had even the slightest bit of intelligence and knew how to mass their forces. They always have more numbers but weaker units who rush everything in their target radius with the main goal being to kill just one of your units. (This strategy is pretty viable if they’re aware they’re in a video game and know the player most likely reset if Donnell ever gets killed.)

The result then with Fire Emblem is that battles aren’t really Lee vs Grant duels of master strategy, but more a puzzle of sorting how you fight through the enemy army with no losses. It’s a lot like Into the Breach on a larger scale.


I was actually thinking about game AI yesterday because I was impressed by the AI in… um… Quantum Break. Honestly though, maybe I just hadn’t played a 3rd person cover shooter in a bit but it seemed reactive and clever in ways many other games aren’t.

Also played some Forza Horizon 3 yesterday and was impressed by that AI too. Likely because I’ve been playing The Crew 2 and the AI in that game is rough in how it cheats around turns and blinking forward when behind (especially the championship races). But in Horizon, I have picked up on the different behaviors of some of the different drivers and it’s cool to even be able to see that. The festival challenges where you have the train or whatever aside.

Those aside, I love many turn-based and tactics and strategy games, but I’m pretty bad at them until I’m tons of hours into the game, so usually it’s not like I can manipulate the AI on that high a level.


I love the game dearly but Endless Legend’s AI is passive to the point that it’s occasionally lore-breaking. I’ve had multiple games where I shared a border with the supposedly hyper-aggressive bug-monster empire and been at total peace with them for much of the game, it’s odd. Certainly doesn’t ruin the game or experience or anything like that, but it is funny when all the art and writing around that civ details their efforts to eradicate all other sentient life forms and then just have them chill and quietly focus on internal development instead of expansion.

Civ V on the other hand gets on my nerve sometimes once you realize that certain empires will always react in the same manner, at least from what I’ve noticed. As soon as I see montezuma nearby I know there’s gonna be 9 Jaguar units stomping through my barely developed early game cities to try and bust up my shit. Thank god civ makes it so easy to defend a city with like one measly archer garrisoned there. Wish there was still some element of surprise regarding how the AI reacts to you.


I agree with you about Quantum Break…the AI is actually pretty aggressive about flanking you and always keeping the pressure up. The game actively punishes you for staying hunkered down in the same place for too long, and encourages you to keep using your powers to manipulate the environment/move around quickly to your advantage.

It actually reminds me of what really annoyed me about the Giant Bomb quick look of that game. I usually really hate it when people say "YOU’RE PLAYING THE GAME WRONG!’, because if there is an intended playstyle, if you’re not naturally getting it, that points to the game doing a poor job of teaching. Plus it gets into a slippery slope of ‘Geez, only good players should be covering games’ bullshit. All that being said, it was maddening to me, watching Jeff go around, barely using his time powers, and playing it like a standard cover-based shooter, when it absolutely is not what the game teaches you. You stick to a cover spot in that game for more than a few seconds, you get killed quick.


A classic answer but F.E.A.R. is my top pick (pdf of that GDC talk) for some pretty simple AI that still feels like it does a bit more than you normally get. While we’ve had tastes of that feel from things like some of the enemies in the original Half-Life back in 1998 and even the basic flanking AI that emerged from Unreal (earlier the same year), something about F.E.A.R. just combines so well to create arena combat puzzles that you could jump into several times and still come away seeing something new. It’s more than the AI (even when you know every trick coded into it and all the stuff it just fakes), it’s also the level design. I’d say it might even be mainly the level design (in the same way that a good multiplayer map lives on that whitebox design, the AI stands above because of the spaces it is able to work inside).

One of those launch games that splits opinions but has ultimately been somewhat forgotten five years later, the most recent Killzone actually did a lot to evoke a similar sense of combat arenas and an almost puzzle-like precision around dealing with dangerous and adapting AIs (possibly only when played on the highest difficulty). Again, it’s the spaces where that combat happens that makes the AI feel right, taking breaks from the more linear corridors and HL2-style variety sections to drop you into an area for some more intense combat and sometimes a few waves of enemies.

Something about how the pieces all fit together is unlike a Halo encounter, while taking away the same satisfaction from the combat loop. The loop gives some games an honourable mention but there is a distinction to be teased out of how the reactive AI feels more like a puzzle that you can always solve (in FEAR or KZ:SF) and less like an encounter whose difficulty can range from trivial to impossible.


I think my favorite AI is actually from AI War: Fleet Command. Its asymmetrical so its not quite approaching the same question as most RTS AIs, but they manage to make engaging with the AI very interesting. Thematically the AI doesn’t divert its overwhelming forces against you because its attention is elsewhere outside the galaxy, but as you progress through a match the AI will begin to pay more attention to you, directing more resources towards dealing with you, reacting to the way you build defenses/where you attack, attempting to counter the strategies you employ.

What more there are some 50 or so different AI types which change significantly how the AI plays. Ranging from aggressive types that will react to incursions with force or more passive types that have heavy static defenses and making army compositions to match. On top of that they have difficulty levels from 1-10 and as you increase the difficulty the AI actually learns new tactics to use against you, its really fascinating to experience.


But the only strategy games that I succeed in are the ones with terrible AI! :joy:


This isn’t a strategy game, but the differing levels of AI difficulty in the Pokémon games is really interesting to look at. I was watched the Pokémon Blue/Yellow speedrun from SGDQ where they discussed a few select “smart” AI in the game that will always use super effective moves against your pokemon; however, those moves very well may be non-damaging moves such as Pigeot using Agility (psychic type) against a Nidoking because technically psychic is super effective against poison. In the most recent generation, they would put a trainer on a route that would only fight you if you beat all the other trainers there, they typically would use a special item or move that would typically be considered a “smarter” strategy. That was really cool to see.


The thing about strategy AI is that it doesn’t actually need to be smart, it just needs to do a decent job of convincing you that it’s smart. Much like the FEAR enemy AI, it’s more about giving the impression than actually doing it.

That said, GalCiv 2 is probably the best strategy AI I’ve ever seen.


I think not only is the problem that AI is time-consuming to code and test and get right, it’s also that it becomes very difficult to scale. If you actually put work into good AI, you then have to put much more work into having some way of toning it down so that less-skilled players can actually handle it. And if you have a masterpiece AI with a billion finely tuned variables, well, good luck figuring out which knob to turn that makes the AI easier but not utterly broken.

There is a huge opportunity for machine learning here, and it seems very unlikely to me that we won’t see ML game AI in games in the next few years. You can already see videos of academic experiments where ML agents are able to play a variety of different game genres (though the games in this case are specially constructed for the experiment.) I think it would be feasible to dumb them down enough to provide easier difficulty levels. Though that’s just a hunch.

Anywho - tangentially: I am beginning to think I should try Rise of Nations finally, between this and the recent super deep dive interview with Brian Reynolds on Soren Johnson’s podcast (Designer Notes). It sounds like, at the very least, they had some great ideas going into it. I don’t really know why I missed it. Hell, I played Empire Earth when it came out…


Y’all talking SMAC???

Alpha Centauri was a transformative game for me, because even on lower difficulties you could see distinct playstyles amongst the AI, especially in the case of someone like Chairman Yang.


Oh, that’s another reason to listen to Brian Reynolds’ interview on Designer Notes. They get deep into talking about SMAC. It’s super interesting.

One of the best games, period.