Why AI still sucks in video games — and how DirectX 12 could change it for the better

 

 One of the most common user complaints about gaming is the AI. It’s been this way for decades, which doesn’t seem to make sense — computers today are vastly more powerful than systems built 30 years ago. We recently sat down with Brad Wardell, CEO of Stardock and the author of the AI in Galactic Civilizations II about this problem — and how DirectX 12 and improved multi-threading support offer hope of improvement.
The underpinnings of RPG AI emerged relatively quickly. Games like Ultima IV (1985) offered comprehensive conversation trees that allowed players to query NPCs (non-player characters) on specific topics. A major part of Ultima IV’s gameplay involve traveling the world and learning from its various denizens. The approach feels deep, even today, despite the fact that U4 shipped on two double-sided floppy disks.

We had day/night cycles when these graphics were cutting-edge

Ultima V (1988) added a day / night cycle, which means that two of the fundamental building blocks of RPG gaming — robust NPC conversations and realistic life cycles — were shipping nearly 30 years ago, in games of less than 1MB. That doesn’t mean we’ve made no improvements since, but according to Wardell, AI has actually become worse in some respects. There are several interlocking reasons why this is the case.

Scale, cost, and complexity

The first problem, Wardell explained, is that good AI doesn’t really sell games. Gamers may value good AI, but it’s not the determining factor in whether people buy a title. At the same time, AI is now tightly coupled to graphics and must be communicated visually. It’s no longer enough to tell the player, “The shopkeeper looks frightened,” as players want to see that the shopkeeper is frightened.

Yes, this guard is attempting to arrest a fish

Virtually all of what we call AI in games relies on scripted sequences, either hand-written or procedurally generated. Bethesda’s Radiant AI, for example, uses procedurally generated scripts to create goals and tasks for NPCs, then allows them to determine how best to fulfill their needs. The company famously had to tone down how the system functioned in The Elder Scrolls: Oblivion after it became clear its original approach produced unintended consequences. In one case, NPCs addicted to a drug called skooma would buy the drug until they ran out of money, then kill the dealer and steal his supply. Players would arrive to talk to the dealer as part of a quest chain, and find he was already dead.

Multithreading and the AI-vs-graphics tug of war

Unlike graphics workloads, which can be easily parallelized and scale well, artificial intelligence processing is difficult to multithread. The job-based systems that many games use for multithreading don’t tend to work well for AI processing, Wardell explained, because the AI is constantly asking questions and must receive immediate answers. Job-based systems are designed to launch and process workloads, not facilitate constant communication between the job and the host CPU.
To understand this better, consider interactions between the player and an NPC shop owner. Bethesda games typically offer you the option to steal goods without penalty, provided no one else sees you doing it. This means the game is constantly calculating the shopkeeper’s line of sight. If you’re trying to sneak around a camp full of bandits, the AI is checking to see how much noise you made and whether the bandits heard it. In both cases, the AI is asking a question and receiving an answer many times a second.
A critical part of the reason why AI doesn’t seem to have advanced much, even though computers today are orders of magnitude more powerful than their predecessors, is because the overwhelming majority of that compute capability has been devoted to making games prettier, not smarter.

Dogmeat in FO3 vs. FO4. He responds to your actions and emotes various states in Fallout 4

Because AI actions and reactions are synched to graphics, AI workloads need to be completed in time for an NPC to react to what’s going on in the world around them. That means AI processing needs to happen within a 10-20ms timeframe — and since AI workloads don’t thread well, the AI is probably running on a CPU core that’s also handling other tasks. Most games display more than one NPC at a time, so all of their collective AI calculations must be handled simultaneously.

The impact of audio

Audio processing doesn’t impact AI calculations the way graphics do, but players now expect titles to be fully voiced. This has its own impact on game design. In The Elder Scrolls: Skyrim, the phrase “I used to be an adventurer like you…” became a popular meme partly because it captured the monotony of traveling from town to town and hearing the exact same banal observation repeated over, and over, and over.
I was curious on this point, so I ran a completely unscientific survey of my friends who had also played Skyrim. The question was: “If you had to guess, how many unique lines of ambient conversation dialog do you think a generic Skyrim guard has to say?”

Oh god. Not again.

The highest guess I got was 75. Most people, including me, guessed between 10 and 30. The actual number is well above 400, even after discarding near duplicates. But since not every type of guard can make every remark, and some remarks only trigger under specific conditions, NPC guards still feel like paper cutouts, rather than actual persons. Guards may say hundreds of things, but we remember the handful they repeated ad nauseam.

Stealth mechanics are often frustrating

Before the advent of voice acting, a small team of writers could theoretically churn out a huge number of responses. Today, all those parts are voice-acted. If dialog writing scaled poorly 20 years ago, dialog voice-acting scales even more poorly today. Keep in mind, we’re still just talking about the generic guards. If a named NPC has unique responses to your class, gender, or other attributes, all of them must be recorded. This is why many games tend to use a small pool of voice actors, and it’s part of why dialog trees are often narrow, with multiple lines of conversation offering only minimal differences between them.

How Ashes of the Singularity and DX12 can change things

Up to now, we’ve discussed RPG AI, because it’s one area where poor AI sticks out the worst. Wardell’s expertise is in RTS and strategy AI, however, and with Ashes of the Singularity already in Early Access, we wanted to talk about how RTS AI has evolved since the days of Starcraft and the original Command and Conquer, and where it might be headed in the future.
One thing that sets Ashes apart from most games on the market today is that Oxide is deliberately choosing to require a quad-core CPU to run the game. This effectively limits Ashes to the slice of the market that owns a quad-core (the Steam hardware survey shows the market is split on this point, with roughly half of users owning a dual-core system and half with a quad-core).
The reason Oxide opted to require a quad-core CPU is partly because it wanted to improve AI. According to Wardell, the basic structure of AI in any RTS to-date looks like this:
while(true)
{
DoStuff();
DoAI();
DrawStuff();
}
Here’s Wardell:

“What Ashes does is eliminate the game loop thing entirely. But this comes with a heavy up-front cost. It’s expensive (CPU wise) to have an asynchronous job system that can answer questions as a job. That cost comes in the form of us having to require four CPU cores as a minimum requirement.

“What we get in return, however, is that the AI is asynchronous to the rest of the gameplay so we can ask lots and lots of questions (i.e. How dangerous is this threat? What is their firing range? What is my firing range? What terrain is nearby? What allies do I have nearby? Can I ask a nearby unit for help? etc.). And have these answers be provided really fast because we know the user has at least four CPU cores so we can make certain assumptions on response time.

“Eventually, every game will require four CPU cores because of what I’ve outlined here. Being able to have your job system (which is, by definition, asynchronous) be able to answer questions (which is a synchronous task by definition i.e. Q->A) in real-time means you can much more sophisticated AIs in real time.”

DirectX 12 isn’t directly intended to improve AI calculations, but it has a positive effect nonetheless. Distributing workloads across multiple cores reduces overall CPU overhead and gives developers more headroom for other kinds of work.

Do players actually want great AI?

As frustrating as poor AI can be, the alternative may be worse. Imagine playing Skyrim and fighting your way through a dungeon, only to discover that an NPC group of bandits had located and raided the treasure first, following the same clues that you did. Enemies in these games often respond hilariously poorly to stealth-based strategies, quickly forgetting that they’re under attack and walking mindlessly over the bodies of stealth-killed compatriots. This scenario could obviously be improved — but a game where guards established realistic choke points and immediately ordered additional patrols might not be much fun, particularly for low-level players.

Things like this could happen less. We wouldn’t mind.

Games like Dishonored, which implements stealth tactics far more effectively than Skyrim or Fallout, is also a very different title. Levels are self-contained, with far fewer NPCs and no sprawling open-world. These trade-offs are not accidental.
Are there players who would relish the chaos of an open-world game in which random bandits could depose rulers and dragons would loot dungeons before you ever reached them? Undoubtedly. I suspect some of them already play Dwarf Fortress. But this kind of freedom makes it impossible for a game to present the player with a coherent narrative. These are the kinds of trade-offs that we won’t see resolved any time soon, since they strike at the inherent difficulty of creating strong narratives while allowing for interactive gameplay rather than any kind of limited resource.
The advent of DirectX 12, however, should give developers more freedom to experiment with better AI simply by giving game engines the option to utilize multi-threading more efficiently. We’re hoping this yields improvements across the entire game spectrum. We’ve been stuck with wooden cutouts and thin interaction for a long time now.