(Cross-posted from my tumblr.)

Artificial intelligence is an area I think about a lot. My undergrad research was primarily concerned with intelligently coordinating teams of agents in games and now, a lot of the work at my day job involves researching new ways to guide AI agents to intelligently reason about real-world problems in simulations.

In a simulation, the more intelligent an agent is, the better the simulation. Although they have a lot in common, games are very different from simulations. The goal of a simulation is training, whereas the goal of a game is (arguably) entertainment.

I’m going to talk a little bit about what this distinction means for the design of enemies in games.

Gamers often talk about wanting “more realistic” or “more intelligent” enemy AI in games, but I don’t think that that’s necessarily desirable. For one, players never notice AI in games unless it’s bad. When working on my graduation project, we poured untold hours into researching and designing a group coordinator AI that could dynamically generate and employ strategies, but almost no one who played the game ever commented on (or seemed to observe) the intelligence of the AI.

The party member AI actually did have to be turned down in early testing because it was way too intelligent and it made the player feel like an accessory. The party AI liked to employ this particularly wicked strategy:

Have the party healer cast every status augmentation on the grenadier
Have the primary attacker run in ahead of everyone else and start taunting and luring all the enemies into one place
Have the grenadier constantly spam every area-of-effect attack on the now-gathered enemies
Given the rules of the game, this was the optimal strategy. It was brutally effective, but it wasn’t very fun. For one, it was repetitive. For two, there was never a reason to deviate from it. We decided to tweak the party member AI so that, rather than trying to win the battle, its primary goal would be to stay near the player, support them, and make a show of hitting things close to the player.

As a result, even though the AI was much stupider, it appeared more intelligent because you noticed it more, and you always saw it doing things that helped you, the player. They were strictly worse as teammates but made the game more fun to play.

Enemies have the same problem.

Typically, when a player gets outsmarted by an enemy, their response isn’t “wow, that was clever!” The response is usually frustration. “What was I supposed to do?!” “The game is cheating!” Even if the enemy is playing by the same rules as the player, you don’t want to admit to being outsmarted by a machine, and so your natural response is to blame the game for being bad.

While it’s all well and good to say “well, that’s the player’s fault, not the game’s,” it’s a lot easier to change a game than it is to change human nature. So it’s a good idea to keep this in mind when designing AI for enemies in our games.

When designing the encounters in Vacant Sky Awakening, Kate and I sat down and decided upon a unique theme or challenge, that each encounter would present to the player. The enemies in the encounter are designed to act according to a predictable set of rules that the player can learn, and the challenge comes from figuring out a way to overcome the challenge with the resources at your disposal.

It is not an easy game. Most of the encounters are designed to beat you. But each encounter has multiple solutions depending on your skill and party composition, and the idea is to let the player express creativity in solving the problems the encounters pose. Because so much of what the player can do in battle is a consequence of their earlier actions in the story, forcing the player to take ownership of their decisions and use what they’ve left themselves with in order to solve problems gives them a sense of responsibility, which will hopefully drive personal engagement with the game.

Making the enemy AI realistically intelligent would be easy: have all of the enemies gang up on the strongest party member, kill them ASAP, then move onto the next strongest. That’s how most people play RPGs. But being on the receiving end of that isn’t fun. It’s frustrating and it feels unfair.

Instead, design enemies like puzzles, easy to understand but challenging to conquer. Games aren’t simulations, and in areas like artificial intelligence, realism can be the enemy of fun.