• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
  • The Politics forum has been nuked. Please do not bring political discussion to the rest of the site, or you will be removed. Thanks.

Age of empires IV to implement 'Merciless' A.I. that will learn from players, eventually becoming unbeatable

cormack12

Gold Member
Mar 21, 2013
10,739
28,478
1,550
Source: https://www.kotaku.com.au/2021/08/age-of-empires-4-ai-machine-learning-difficulty/

How long will it be before RTS games have an AI so good that it defines how humans play the game?

In a new interview with World’s Edge franchise creative director Adam Isgreen and Relic game director Quinn Duffy, I asked the pair about an age-old annoyance: AI. When you up the difficulty to get a real challenge, the dev answer has simply been to let the AI cheat — or give them absurd buffs to resources and other tricks to paper over the cracks of the AI not being smart enough.

Age of Empires 4, gracefully, won’t resort to either of those old-school methods. But Isgreen and Duffy both revealed that Age of Empires 4 has been toying around the idea of a Merciless AI. It’s inspired by some of the work that went into Killer Instinct’s Shadow AI, which learnt from the moves players made.

Age 4 already used machine learning to train its AI, but after the game launches this October, the developers will probably add this harder difficulty mode that will just continually learn from whatever the current meat is. “We use machine learning in training the AI right now, but we want to take that even further. Down the road, not at launch, we’ll probably look into having a merciless AI that keeps learning the more people play against it, to the point that it’ll be unbeatable,” Isgreen said. “But we’re OK with that, because if you opt into that difficulty, if you want to opt into making the AI better at beating you, go right ahead.”

“We can use that in different aspects in terms of how we cut it up for the different difficulties. Like, we did a lot of that in Killer Instinct 2, we had AIs that would learn how to play like players down to even the taunts, you know for a game like KI, it’s like when did [the player] teabag, how did they move around while they were waiting for a player to get up,” Isgreen added. “I think we can start applying a lot of that to a real time strategy game. But the AI it does things that players do, we try to imitate and add in strategies that players use.”

For lower difficulties, however, Isgreen emphasised that the AI is tuned to create exciting matches for players. He recalled a GDC talk on real-time strategy balancing years ago that was run by industry heavyweights including Chris Taylor. “I was like: AI should only cheat to make the game experience more enjoyable for the player, and I got booed by people,” he said.



“The guy from Total Annihilation was there — Chris Taylor was running the thing but one of the other guys from Total Annihilation was like, ‘AI should be perfect, it should never cheat’, and everybody was applauding.”

“And I was like no, you don’t get it: we make entertainment. I don’t want AI to cheat in terms of getting buffs and bonuses — that isn’t satisfying for me. But when difficulties are lower, it’s important to realise that drama is almost more important than perfect balance. You want highs and lows emotionally. An AI that can do that and generate a fun fight that people walk away from [saying], ‘Wow that was great, I just pulled it off’ — even if that was fudged a little bit, that entertainment is so valuable. But as you turn up the difficulty, that’s when all of that stuff diminishes.”