It’s been a while since I looked at how Valve does it but it could be called a primitive expert system. And while the HL1 grunts were extraordinary for their time, HL2’s combine grunts are still pretty much the gold standard. Without the AI leaking information to the player via radio chatter it would feel very much like the AI is cheating because yes, HL2’s grunts are better at tactics than 99.99% of humans. It also helps that you’re a bullet sponge so them outsmarting you, like leading you into an ambush, doesn’t necessarily mean that you’re done for.
OTOH they’re a couple of pages of state-machines that would have no idea what to do in the real world.
Also, for the record: “AI” in gamedev basically means “autonomous agent in the game world not controlled by the player”. A “follow the ball” algorithm (hardly can be called that) playing pong against you is AI in that sense. Machine learning approaches are quite rare, and if then you’d use something like NEAT, not the gazillion-parameter neutral nets used for LLMs and diffusion models. If you tell NEAT to, say, drive a virtual car it’ll spit out a network with a couple of neurons, and be very good at doing that but be useless for anything else but that doesn’t matter you have an enemy AI for your racer. Which probably is even too good, again, so you have to nerf it.
I still remember fighting grunts in the original half life for the first time and being blown away. Your work makes games great!
It’s been a while since I looked at how Valve does it but it could be called a primitive expert system. And while the HL1 grunts were extraordinary for their time, HL2’s combine grunts are still pretty much the gold standard. Without the AI leaking information to the player via radio chatter it would feel very much like the AI is cheating because yes, HL2’s grunts are better at tactics than 99.99% of humans. It also helps that you’re a bullet sponge so them outsmarting you, like leading you into an ambush, doesn’t necessarily mean that you’re done for.
OTOH they’re a couple of pages of state-machines that would have no idea what to do in the real world.
Also, for the record: “AI” in gamedev basically means “autonomous agent in the game world not controlled by the player”. A “follow the ball” algorithm (hardly can be called that) playing pong against you is AI in that sense. Machine learning approaches are quite rare, and if then you’d use something like NEAT, not the gazillion-parameter neutral nets used for LLMs and diffusion models. If you tell NEAT to, say, drive a virtual car it’ll spit out a network with a couple of neurons, and be very good at doing that but be useless for anything else but that doesn’t matter you have an enemy AI for your racer. Which probably is even too good, again, so you have to nerf it.
TIL OTOH is short for On the Other Hand
Is it? I always thought it was Off The Top Of my Head for some reason.
Could be, idunno, I just googled it.
<sweats in Opposing Force>