4 ms·

Firaxis’s team is quite skilled, and many of its designers also code, which helps them build features more intentionally than at many other studios. I’m pretty

by caseyy 2y ago

Firaxis’s team is quite skilled, and many of its designers also code, which helps them build features more intentionally than at many other studios. I’m pretty sure they considered doing more autonomous AI for Civ VII, but there are really good reasons not to. In most strategy games, rule-based AI is easier to author, contextualize, and theme. Players usually prefer AI agents with memorable and varied traits instead of excellent high autonomy (complex utility, HTN, ML) AI. They like evident strategy more than complex strategy. Smarter and harder AI sounds appealing on paper, but it’s often less fun for most players who want a predictable, beatable opponent. High-fidelity human-level opponent AI is only really appreciated in eSports, where it serves a training purpose. However, eSports people have fun in an aim trainer—they are not the “normal” gamer. Most people want “really weak chess” and wouldn't enjoy hard or complex chess. They like game AI opponents acting in simple and vulnerable patterns they can outsmart and feel clever. And the more they dominate the agents by learning them, the more they like the game. Even if they say they don't, the user testing shows they do. The cheat bonuses that are super easy to win against are probably intentional :) There is also another reason for building simple game AI — the illusion of intelligence is not created by complexity but rather by anthropomorphization/personification. It's the same as with LLMs — the big breakthrough was when they started speaking like humans, and all further complications aren't adding much to the broad and intuitive feeling of their intelligence. Similarly, a game AI agent that tells you what they are thinking achieves more perceived intelligence than an AI agent that is actually more intelligent. With this in mind, game developers don't waste much time on deep intelligence. This time is better used making agents bark, animate, display human reactions, show context awareness, etc. It is very much about the agents being showhorses, not workhorses. I say this from pretty long AAA game AI design and architecting experience. It’s wiser to do what Firaxis did.

somenameforme 2y ago
It's not like the choices are between a 'lively' cheating AI that can't play worth a fiddle without massive cheat bonuses, and a soulless unbeatable superhuman AI that plays fair. If it were possible to create a superhuman fair playing AI then the Civ bonuses themselves would create personality (in fact often strongly related to the desired personality characteristics of each civ!) even without additional 'hard coded' imperatives, which would of course also be possible. And in this scenario create a compelling weaker AI would be much easier than trying to create a compelling strong AI (from a very weak base). There will need to be some 'move' evaluator function and simply picking weaker 'moves' would so fine. If that creates uncompelling AI then we could get back to cheating and impose resource penalties, but with much more compelling play. E.g. - kind of like playing a strong chess engine with knight odds.
- caseyy 2y ago
  When one builds game AI, one of the core design decisions is whether they want the system to be autonomous or authorable. You are right, its not a strict dichotomy. But both choices offer many advantages that an undecided/mixed approach doesn’t. Look up different game AI architectures: scripted, rule-based, finite state machines, behavior/decision tree, utility, planning, hierarchical task networks, and machine learning. Hard-coded imperatives are very difficult on highly autonomous architectures, as is complex systemic behavior in authored ones. Consider how hard it is to make an autonomous AI system like ChatGPT implement one simple hard imperative: not revealing the system prompt. Weak AI and strong AI are ambiguous terms. Most junior-level software engineers can create very complex gameplay systems if you're talking about complexity. It’s much harder to build simple ones that effectively create the illusion of intelligence on limited hardware. If you are talking about difficulty, then challenging AI is no harder to make than easy AI. Usually, game designers will naturally create hard AI as they play the game so much they become experts, and player testing is used to make it easy again. As you describe it, the move evaluator is a utility function approach that is neither very authorable/personable nor very autonomous. It is pretty limiting on both fronts, so it’s rarely used. A simple rule-based approach is more straightforward to balance and personalize for each agent/archetype. And difficulty scaling or balancing the game differently doesn’t require what you propose. They just chose not to do the balancing you want. There’s nothing wrong if you prefer a utility approach. We all have our preferences, and I like some pretty clever AI systems like GOAP that aren’t always a good fit for games like Civ. They are cool and smart systems but players rarely respond better to them than simple FSMs or rule-based game AI, which are easier to author. The utility approach is somewhere in the middle and could also be made context-specific with enough separate utility functions the agent would evaluate. But it is so much more work than to define simple if-then rules. :)
  - somenameforme 2y ago
    I find the detail interesting but I think this is somewhat tangential. What we're talking about is cheating vs non-cheating AI. How it happens to be implemented is largely inconsequential. And it's obviously entirely possible to create a non-cheating AI with scalable difficulty, and there's not a player alive that'd miss the current systems of incapable AI that generates a challenge by excessive cheating.
ProjectArcturis 2y ago
I would just like an AI that isn't so incompetent that it breaks immersion in the game. In Civ 6, on Deity, if you survive the first 50 turns, when the AI had a massive cheat advantage, it is virtually impossible to lose to the AI militarily. In later stages they essentially never use planes, or any kind of coherent tactics. The AI certainly does tell you what it's thinking: "Duuuuhhhhhhhhhhh"

by caseyy 2y ago

somenameforme 2y ago

It's not like the choices are between a 'lively' cheating AI that can't play worth a fiddle without massive cheat bonuses, and a soulless unbeatable superhuman AI that plays fair. If it were possible to create a superhuman fair playing AI then the Civ bonuses themselves would create personality (in fact often strongly related to the desired personality characteristics of each civ!) even without additional 'hard coded' imperatives, which would of course also be possible. And in this scenario create a compelling weaker AI would be much easier than trying to create a compelling strong AI (from a very weak base). There will need to be some 'move' evaluator function and simply picking weaker 'moves' would so fine. If that creates uncompelling AI then we could get back to cheating and impose resource penalties, but with much more compelling play. E.g. - kind of like playing a strong chess engine with knight odds.

caseyy 2y ago
When one builds game AI, one of the core design decisions is whether they want the system to be autonomous or authorable. You are right, its not a strict dichotomy. But both choices offer many advantages that an undecided/mixed approach doesn’t. Look up different game AI architectures: scripted, rule-based, finite state machines, behavior/decision tree, utility, planning, hierarchical task networks, and machine learning. Hard-coded imperatives are very difficult on highly autonomous architectures, as is complex systemic behavior in authored ones. Consider how hard it is to make an autonomous AI system like ChatGPT implement one simple hard imperative: not revealing the system prompt. Weak AI and strong AI are ambiguous terms. Most junior-level software engineers can create very complex gameplay systems if you're talking about complexity. It’s much harder to build simple ones that effectively create the illusion of intelligence on limited hardware. If you are talking about difficulty, then challenging AI is no harder to make than easy AI. Usually, game designers will naturally create hard AI as they play the game so much they become experts, and player testing is used to make it easy again. As you describe it, the move evaluator is a utility function approach that is neither very authorable/personable nor very autonomous. It is pretty limiting on both fronts, so it’s rarely used. A simple rule-based approach is more straightforward to balance and personalize for each agent/archetype. And difficulty scaling or balancing the game differently doesn’t require what you propose. They just chose not to do the balancing you want. There’s nothing wrong if you prefer a utility approach. We all have our preferences, and I like some pretty clever AI systems like GOAP that aren’t always a good fit for games like Civ. They are cool and smart systems but players rarely respond better to them than simple FSMs or rule-based game AI, which are easier to author. The utility approach is somewhere in the middle and could also be made context-specific with enough separate utility functions the agent would evaluate. But it is so much more work than to define simple if-then rules. :)
- somenameforme 2y ago
  I find the detail interesting but I think this is somewhat tangential. What we're talking about is cheating vs non-cheating AI. How it happens to be implemented is largely inconsequential. And it's obviously entirely possible to create a non-cheating AI with scalable difficulty, and there's not a player alive that'd miss the current systems of incapable AI that generates a challenge by excessive cheating.

ProjectArcturis 2y ago

I would just like an AI that isn't so incompetent that it breaks immersion in the game. In Civ 6, on Deity, if you survive the first 50 turns, when the AI had a massive cheat advantage, it is virtually impossible to lose to the AI militarily. In later stages they essentially never use planes, or any kind of coherent tactics. The AI certainly does tell you what it's thinking: "Duuuuhhhhhhhhhhh"