How Could Hybrid AI Enhance Gameplay Dynamics and Decision-Making in Civilization 7?

[steveanderson]

Hey everyone,

I've been thinking about the potential impact of Moderator Action: *SNIP* Ad links not allowed. -lymond on gameplay in Civilization 7, and I’d love to hear your thoughts!

How do you think Hybrid AI could enhance gameplay dynamics and decision-making for both players and AI opponents? Here are a few points to consider:

1. Adaptive Strategies: Could Hybrid AI analyze player strategies in real-time and adapt its tactics accordingly? How might this affect long-term gameplay?
2. Improved Diplomacy: What role do you think Hybrid AI could play in making diplomatic interactions more nuanced and realistic? Would it lead to more engaging alliances and rivalries?
3. Dynamic Difficulty: How might a Hybrid AI approach help balance difficulty levels, ensuring that players feel challenged without being overwhelmed?
4. Player Engagement: Do you believe that an AI capable of learning from player actions would lead to deeper strategic planning and more immersive gameplay?

I’m really curious about how you all envision Hybrid AI shaping the future of Civilization 7. Looking forward to your insights!

Best,
Steve

Moderator Action: Link removed. leif

 

[stealth_nsk]

It's generally a lot of computation power, which instead of improving game would make it worse. In single player people learn to play against AI and value roleplaying from AI. Super-optimized learning treacherous opponents are expected in multiplayer, not as AI.

 

[1SDANi]

It's generally a lot of computation power, which instead of improving game would make it worse. In single player people learn to play against AI and value roleplaying from AI. Super-optimized learning treacherous opponents are expected in multiplayer, not as AI.

Didn't the original Crash Bandicoot have dynamic difficulty?

 

[stealth_nsk]

Didn't the original Crash Bandicoot have dynamic difficulty?

I think it's a bit different kind of game

 

[1SDANi]

I think it's a bit different kind of game

It was a game on the original Playstation. Don't really see what's computationally expensive about having AI switch between two or three agendas depending on game state, it's basic state machine type stuff that civ has been using since the first game.

 

[stealth_nsk]

It was a game on the original Playstation. Don't really see what's computationally expensive about having AI switch between two or three agendas depending on game state, it's basic state machine type stuff that civ has been using since the first game.

Dynamic difficulty is not the same thing as self-learning AI. Dynamic difficulty surely could be done, but I see no reason for using it in Civ, since we have manual difficulty.

 

[1SDANi]

Dynamic difficulty is not the same thing as self-learning AI. Dynamic difficulty surely could be done, but I see no reason for using it in Civ, since we have manual difficulty.

In theory you're right, in practice they might as well be the same. Using machine learning for dynamic difficulty, adaptive behaviour, and improved diplomacy is like using a nuke to kill a spearman. None of what OP describes requires machine learning.

 

[stealth_nsk]

In theory you're right, in practice they might as well be the same. Using machine learning for dynamic difficulty, adaptive behaviour, and improved diplomacy is like using a nuke to kill a spearman. None of what OP describes requires machine learning.

It depends on the exact definition of machine learning. I worked in this field with tools ranging from very simple regression models to very complex neural networks with transformer architecture, and all of them were called "machine learning". Simple machine learning models could actually more or less fit things like "adaptive behavior", but it's more a question of wrong goals than wrong tools. Having stable behavior allows developers to test and tune AI, providing great player experience in the end, while auto-tuning AI behavior in strategic games looks like a recipe for failure.

 

[1SDANi]

It depends on the exact definition of machine learning. I worked in this field with tools ranging from very simple regression models to very complex neural networks with transformer architecture, and all of them were called "machine learning". Simple machine learning models could actually more or less fit things like "adaptive behavior", but it's more a question of wrong goals than wrong tools. Having stable behavior allows developers to test and tune AI, providing great player experience in the end, while auto-tuning AI behavior in strategic games looks like a recipe for failure.

Metal Gear Solid V had enemies field units to counter the strategies you've been using, to use civ terms, if you're using a lot of swordsmen, AI that want to invade you would field units that do well against swordsmen. That is a form of adaptive behaviour that is no less stable than anything civ already does. In fact, I'd be surprised if civ doesn't already employ it on some level.

 

[stealth_nsk]

Metal Gear Solid V had enemies field units to counter the strategies you've been using, to use civ terms, if you're using a lot of swordsmen, AI that want to invade you would field units that do well against swordsmen. That is a form of adaptive difficulty that results in behaviour no less stable than anything civ already does. In fact, I'd be surprised if civ doesn't already employ it on some level.

It's not self-learning AI, it's AI reacting to player actions. Any game AI does this in some form or another. It's generally the same as civ AI attacking you if you don't have enough troops.

 

[1SDANi]

It's not self-learning AI, it's AI reacting to player actions. Any game AI does this in some form or another. It's generally the same as civ AI attacking you if you don't have enough troops.

That's my entire point. Anything self-learning AI can be reasonably used for in civ can be better implemented with less overkill methods. People all too often use newfangled machine learning buzzwords to describe things Crash Bandicoot did in the 90s with a fraction of the processing power.

 

[Vitruvian Guar]

After the reality check of playing Old World, I started suspecting that maybe mediocre AI in civ games is actually an intended feature, not a bug.

When AI is good enough to create some semblance of competition while having enourmous bonuses it can still be fun to win against it. Additionally you get the feeling of superiority for overpowering the stupid machine even despite it's head start.

When AI can kick your ass, even without special bonuses... that's not fun at all - its humiliating. It alienates the casual player and for a hardcore one it takes extra effort to start having fun from the game.

 

[stealth_nsk]

After the reality check of playing Old World, I started suspecting that maybe mediocre AI in civ games is actually an intended feature, not a bug.

When AI is good enough to create some semblance of competition while having enourmous bonuses it can still be fun to win against it. Additionally you get the feeling of superiority for overpowering the stupid machine even despite it's head start.

When AI can kick your ass, even without special bonuses... that's not fun at all - its humiliating. It alienates the casual player and for a hardcore one it takes extra effort to start having fun from the game.

The big problem with mediocre AI in Civ is what playing against it is often a very uneven experience. If you start near AI and get early war, AI could crush you just by a sheer number of troops. But if it lets you develop, you could totally dominate the late game.

However, I'm not sure the solution should be in improving AI. Maybe instead the game should change how AI bonuses are applied - and Civ7 ages look like a perfect base for it.

 

[Vitruvian Guar]

I think it would solve it. A smarter AI could have fewer bonuses so it wouldn't curb stomp you at the start of the game because it's not stronger than you at that point. IIRC from my Civ5 modding days, the AI did use to have scaling bonuses that increased with age. Those just didn't matter because they never outcompeted snowballing. Now if the AI is too smart and without any bonuses it still beats a player easily, then dumb it out. It's much easier to make a smart thing stupid than vice versa.

The chess world has had the "problem" of the AI being too good for almost 3 decades now. Yet you can go online now and play an interesting game against of chess against a computer, as evenly matched or not as you want. But its definitely true that if the demand for smart chess AI isn't here, FXS isn't going to invest in making a really good one only for 99% of players to dumb it down. I just don't think previous Civ AIs were anywhere near that level - especially Civ5 and 6.

I think Rise and Fall completely ruined the AI in Civ 6 and it never recovered since then. Especially with the optional golden/dark age mechanics - some civs just loose all their cities even of Deity and you end up playing against barbarians.

I'm not sure how much the situation with chess is comparable. There you can easily regulate the competence of the AI by limiting the depth of the search. I don't think there is an analogy for that in civilization. But yes, probably, dumbing down a competent AI is easier than designing one in the first place. Sadly, FRX isn't really motivated to do either.

 

[Vitruvian Guar]

The big problem with mediocre AI in Civ is what playing against it is often a very uneven experience. If you start near AI and get early war, AI could crush you just by a sheer number of troops. But if it lets you develop, you could totally dominate the late game.

One of my best Diety games was when an AI waged early war against me. Focusing on defence from turn one allows to meatgrind all the units the AI sends, while leveling up yours and then you can change the tide and grab enemy cities. After that the game is a cake walk but the initial first war is very exciting. Though, I don't think I would be able to do it without the early governor title from the secret societies optional rule set.

 

[ggalindo001]

I think many have highlighted the true challenge with any AI -- the more "systems" introduced, the harder it is for any algorithm to make "instinctive" choices that are beneficial in the long run.

I do think that the Civ VII devs are looking at certain gameplay elements with AI enhancement in mind (certainly the commanders and able to route as one stack and deploy 1UPT will help tremendously). Also, breaking into hard ages help with things that are hard to manage (optimal upgrade paths, steamrolling based on optimizing in the earlier game, etc.)

I recall either Civ V or Civ VI boasted having multiple layers of AI -- tactical and strategic. Unfortunately it was like 2 different systems passing in the night.

I also agree that a competent AI (like Old World) can be very off-setting for newer players and cause players to just simply stop. Also, Old World AI is still aligned in a relatively basic game mode (the old-world, so no air, etc.)

What I wish for is an AI "tuning" ability -- where I can through sliders or other means, amp up or down things like AI War Aggressiveness, AI Diplomacy, Tall v Wide, Over or under index native playing style for the civ, levels of "diplomatic blocs", etc.

I do wish that the Civ VII devs would look at AI similar to Old World in terms of "competency" -- and then "dumb it down" for the lower difficulty settings.

And not take the road of difficulties just simply around starting (or age starting) bonuses.

What I could see is Diety -- 95% always Max/Min (meaning that 19 out of 20 decisions will always be optimal) -- Emperor -- 80% always Max/Min (4 out of every 5 decisions always optimal) -- Prince -- 75% always Max/Min, etc. -- with lower levels closer to the 50% range.

This could give great challenge to higher level players -- but introduce potential mistakes that allow for a player to also exploit if seen.

That, on top of sliders that allow for tuning of playstyles for the AI (or even sub-components of AI) would be great.

For example, even on the lowest difficulty levels, starting against Rome should feel very different than starting against Cree (I know they won't be antiquity civ) -- but on higher difficulty levels, you get punished if you don't prepare immediately against Rome (or punished differently against Cree).

I'm somewhat optimistic about Civ VII AI -- with no basis for this optimism other than how the game is structured. And for the fact that in many cases, the choices taken away from the player (like workers for example), make it EASIER for the AI to "automate" the same way.

Let's just hope we don't have seafaring Civs in the exploration age continuing not build ships, harbors, etc. Or worse, build navies that cannot ever get to oceans.

 

[rocksinmypath]

Player Engagement: Do you believe that an AI capable of learning from player actions would lead to deeper strategic planning and more immersive gameplay?

It depends. If what you expect is streamlined multiplayer experience, yes. If what you expect is history-themed empire building, probably not. I treat the Civ AI as an environmental element, similar to how I see terrain. I don't want my tile yields to adjust dynamically to my skill level and to how I'm playing the game. I expect the same from the AI. "Hello, I'm John Curtin. I hate wars, but I see you have an unguarded settler, and I've seen you and other players declare surprise wars in these situations to steal the settler, so I'm doing the same to you." That's never going to work for me. It'd completely invalidate so many of the things I love about Civ.

 

[aieeegrunt]

It was a game on the original Playstation. Don't really see what's computationally expensive about having AI switch between two or three agendas depending on game state, it's basic state machine type stuff that civ has been using since the first game.

Hell there are AI mods for Civ6 that do that now, I believe Real AI does, and Roman Holidays’s AI mod is even more sophisticated

 

[ThunderLizard2]

Hopefully higher difficultly levels of AI will be a smarter AI not just an AI with massive bonuses