Do you want an AI that can make reasoned choices?

[Hep Roc Heretic]

Civ IV seems to be the standard by which all other versions are measured. I have played every version of Civ except for IV. Darn, it looks like I missed a good one, eh?

 

[Leyrann]

Civ IV seems to be the standard by which all other versions are measured. I have played every version of Civ except for IV. Darn, it looks like I missed a good one, eh?

I can definitely recommend Civ IV. It's available for a pretty low price nowadays, even more so if you wait for a sale, and while I personally prefer Civ VI, I can definitely see why some people would rather have IV. The AI is much better (I can beat Civ VI on Deity, I'd get completely obliterated in Civ IV even if I'd go several difficulties lower than that), there is a lot more complexity in some game systems such as the technology tree (bring back prerequisites that are not connected by a line!), and it simulates a surprisingly large number of features, such as pop nationality (not very worked out, but it exists), passive trade, et cetera.

Where it, in my opinion, lacks compared to Civ VI is playing the map, city uniqueness, and a ton of flavor stuff. Think traders building roads, district-induced city sprawl, a full civic tree, city states, and so on. I also prefer Civ VI's combat system, though taking another step or two towards more units per tile (similar to the steps taken from V to VI) wouldn't hurt.

I'd say Civ IV is better for the tryharding, optimizing player, but Civ VI gives better general game enjoyment and higher replayability. Also, civilization uniqueness has come a long way.

Civ V, in my opinion, does not match up to either game. I haven't touched it since Civ VI was released (unlike Civ IV, which I picked up around the same time and actually put some two hundred hours or something into).

 

[Hep Roc Heretic]

Civ AI would have to do one or more of the following in order to pass a Turing Test:

1. Send the wrong unit from A to B.
2. Fortifying or deploying a unit and then forgetting about it.
3. Sending a unit on a multi-turn movement and forgetting why you sent it there once it arrives.

We have all done these things. You know it. A computer that makes no mistakes is not human.

 

[steveg700]

I want the AI to at least be able to do things computers do well: compute.

You see civ's with commercial hubs NOT built next to rivers, that's a computer not computing. Or not settling properly at fresh water sources, or prioritizing mountain ranges. For that matter, they should be able to figure out how to combine aqueducts and dams to maximize IZ's. Too many districts with no adjacency bonuses.

I'm in a game right now where I'm in the renaissance era, and the AI civ's still haven't much effort to improve their lands. They all either have one or zero luxuries. This is a pain-in-the-neck, because a player needs trading partners sometimes. And then they settle cities too close to me and then they bother to even send a governor. And again, the unwillingness to get those luxes is hurting them. The AI should be able to compute when it's committing settler suicide.

 

[Grassland Farm]

You see civ's with commercial hubs NOT built next to rivers, that's a computer not computing. Or not settling properly at fresh water sources, or prioritizing mountain ranges. For that matter, they should be able to figure out how to combine aqueducts and dams to maximize IZ's. Too many districts with no adjacency bonuses.

On the flip side: it is through observation of my AI neighbours that I learnt on my second game that I could have two nearby cities building districts "towards" each other to create a blob of adjacency bonuses. Granted, I probably would've found that info on a tutorial somewhere.

I do concur with the lack of land improvements, it's absolutely shameful and sometimes goes on for many eras. Not all tiles are worth spending builders on... but come on, these luxury resources are such a deal breaker.

 

[Benevolent Dictator]

I'm in a game right now where I'm in the renaissance era, and the AI civ's still haven't much effort to improve their lands. They all either have one or zero luxuries.

Decent chance any game where the AI is really short on luxuries is due to the Monopolies and Corporations mode but which was never officially patched (and can't be modded on console).

That mode remains broken from Day 1 where the AI won't start improving their own luxuries until after they've unlocked Corporations from the Economics tech.

 

[TheMeInTeam]

Similarly, it was interesting how the engine only partially was able to play Starcraft on a high level, and showed significant weaknesses there when playing 2/3 races.

Later versions of AlphaStar could play all 3 races, and defeated Serral (the closest thing SC2 had to a top player at the time) + numerous other top players. It did this even when its information was limited to the same things players can see and its APM was limited to a value well below top pro APM.

Doable, but not by Firaxis and not for a game like Civ. At least, not for a decade or so.

Firaxis doesn't have the resources, but as long as you could feed it enough training, machine learning AI in Civ 6 would almost certainly be every bit as oppressive/impossible for humans to defeat in Civ 6 as AlphaStar become in StarCraft 2.

Reason is ideally when something can make up its own mind through reasoning, but neural networks are showing that 'AI's' ar not reasoning, they are learning optimum paths without adequate bias.

If you're looking for generalized AI that "thinks" in the sense that we conceptualize reasoning, that is a much higher standard than we'd need to make an insurmountable opponent in Civ 6, and has different pitfalls.

However, regular machine learning tech available today can handle Civ 6 in principle, best I can tell. While there are many decisions and interacting mechanics, Civ 6 is nevertheless a system with fixed rules, finite possible choices (many of which can be trivially discarded), and defined victory conditions. On a given turn, you only have so many possible moves/actions available to each unit, and only so many possible choices with what to do with cities/citizens/etc. Evaluating all possible combinations of these to select the best ones by brute force is not tractable for humans, but for an AI with many millions of games in training data?

So I would bet that without "reasoning" as you define it, ML could indeed brute force an algorithm that neither you nor any other human alive could reliably defeat, even with no bonuses. For the same reason you can't be Stockfish or AlphaZero in chess.

 

[kaspergm]

I want the AI to at least be able to do things computers do well: compute.

You see civ's with commercial hubs NOT built next to rivers, that's a computer not computing ...

This is obviously true. Question is how much of this is conscious choice by developers to keep AI times down? Personally, I'm sick and tired of the AI almost always settling in exactly the minimum allowed distance from their closest city (only exception seems to be in large desert areas) - when often there will be much better settling locations one or two hexes further away.

 

[Leyrann]

This is obviously true. Question is how much of this is conscious choice by developers to keep AI times down? Personally, I'm sick and tired of the AI almost always settling in exactly the minimum allowed distance from their closest city (only exception seems to be in large desert areas) - when often there will be much better settling locations one or two hexes further away.

If you ask me, even at equal tile quality it's better to settle one tile more than the minimum, simply to have a bit more space between cities once you get all your districts and stuff up and running so that you don't run out of tiles to work. The only exception is if I'm trying to fit in cities in a limited space - if settling a tile further away means settling a city less, I will settle minimum distance.

 

[Heathcliff]

Would be nice if ai had different settings. You could have peaceful ai like it is now. And if you wanted you could have an ai more focused on military.
Even on diety watching lets plays and playing myself ai often has almost zero units. Which is a big difference from other civ and 4x games were AI gets more and more and better and better units as the game progresses.
I understand it is harder to keep cities in civ 6 because of loyalty systems, but ai only war in classical era with warriors and archers and mainly against the player or city states.

Also because the AI doesnt build many units once you have a declared friendship with the ai they will never attack you, in most other 4x if the ai get ahead and have more units they will look for opportunities for war.

The ais problem is not about correct city planning or where to settle it has production bonuses and starting bonuses to make up for that.

 

[Grassland Farm]

Decent chance any game where the AI is really short on luxuries is due to the Monopolies and Corporations mode but which was never officially patched (and can't be modded on console).

That mode remains broken from Day 1 where the AI won't start improving their own luxuries until after they've unlocked Corporations from the Economics tech.

I don't have that game mode (I assume it's in New Frontier, which I don't have) and I still notice an extreme lack of tile improvements on the AI.
I ended my last game with a few conquests to dilute the boredom of a Science Victory and the overwhelming majority of the tiles I conquered were completely devoid of improvements of any kind - no farms, nothing. A few lumbermills, and I think most Irons were improved too - rest was barren.
Actually nice in my case since I just plopped Natural Parks everywhere!

This is making me consider more and more seriously to play Multiplayer and save up some of my evenings in the week for it - if only to fight against someone on roughly equal ground.

 

[AntSou]

I want Firaxis to prioritise designing the game around AI limitations. In Civ VI the AI's inability to handle so many of the game's options ends up having the effect of interactions with AI-controlled civs feeling shallow and ungratifying. On top of that, Civ VI is riddled with little nuisances which could be improved with better AI "cheats". Yes, having a smarter AI is better, but even in the absence of such AI there could be greater care in regards to the handicaps given to the human player.

I'll give two examples from Civ VI to better illustrate my first and second point respectively, in terms of how siege units work in this game. The AI can't competently protect its siege units or appropriately manoeuvre them. Knowing this:

1. Why wasn't the game designed so that Siege Units can stack with other types of military units?
2. Why can't AI siege units move and fire on the same turn starting at a certain difficulty (e.g. Emperor)?

There are three layers of failure in Civ VI in regards to the AI, and all should be addressed to some extent in Civ VII.
Layer 1: actually design a better AI; -> they should probably direct more resources towards AI development.

Layer 2: restrict game mechanics to AI limitations; -> worker charges are a bad idea not because they are inherently so, but because the AI can't quite handle it. They waste those charges, or build it in dangerous spots, and care very little about pillaged improvements. A simpler solution would have been to have workers recharging at the city centre, one charge per turn, before heading out again, rather than disappearing. To compensate, they should probably cost one population, like settlers.

Layer 3: where a mechanic is just too fun to leave out of the game, provide the AI with creative assistances (rather than just boring +X% yields or +1 Unit Y) -> To follow from the example given above, IF workers have charges and they disappear after using them, then the AI worker should have a "zero charge" starting at a certain difficulty. Instead of disappearing after their last charge, they can be kept around in the city centre, one per city, to repair any pillaged tiles (but not create new improvements). Another assistance would be to give AI improvements the effect "Can never be destroyed, only pillaged" by default above a given difficulty.

 

[Leyrann]

2. Why can't AI siege units move and fire on the same turn starting at a certain difficulty (e.g. Emperor)?

While this would indeed achieve the goal of making the AI more of a challenge, I suspect this would in reality be very frustrating to play against for the vast majority of players, even if they support the thought behind it.

A better solution would be to make the AI better at using Logistics, Great Generals and Supply Convoys to achieve the same thing - because if a siege unit has more than 2 movement points, it can move and shoot on the same turn no matter whether or not it has the promotion.

In general, I think it's better to spend more time on improving the AI, and that changing game mechanics for their sake or making their mechanics work differently from the player should be avoided.

 

[Ferocitus]

Disclaimer: I am not an AI expert.
I will be shocked if Civ 7 does not have a neural-net-based AI that is capable of a superhuman level of play.

There is no chance that will happen because computers aren't fast enough and don't have anywhere
near the amount of resources to handle a large enough "neural net" for Civ. And if computers could
do that, you would not like the size of your electricity bill!

Human players can easily "think" many moves ahead in many scenarios and any AI won't be able to
beat them with PCs.

Here's just one very simple example to illustrate what I mean. (There are trillions of others.)
Suppose you have a scout in North America and you spot an opponent's galley headed along the coast
towards your civ in South America. You don't want that civ to find a City State and compete with
you. You know that you can easily buy a coastal tile and stop the galley from moving passed your city
in > 10 turns.

Do you think any AI (now, or for a very long time) is going to be capable of planning that far ahead,
anticipate what you are doing, or might do, and incorporate that into its tactics and overall strategy?
Chess is a far, far simpler game than Civ. For AIs they are not even in the same universe of possibilities.

(I've been using "AI" to optimize ships, wings and for other fluid dynamic applications for > 35 years.
Most (if not all) neural nets and AI algorithms implement some form of internal "optimization".)

 

[Ferocitus]

The main problem is the game engine does not take in all the valid data and make logical conclusions based on it, partially because there is too much data and too many possibilities but there is also a lack of memory by the game engine.
Reason is ideally when something can make up its own mind through reasoning, but neural networks are showing that 'AI's' ar not reasoning, they are learning optimum paths without adequate bias.

They are also extremely brittle. For example, replace one tile's terrain with a different one and
the strategy they adopt could be unrecognizably different.

 

[reddishrecue]

Players playing against the AI could learn from AI if the AI makes good choices.

 

[Chekko]

An AI customizer would be fun, where you customize how the AI should act. I'd like passive Civs to act more realistically and not realize they are in a game while aggressive Civs are more human-like where that late game victory is everything.

 

[Siddorm]

There is no chance that will happen because computers aren't fast enough and don't have anywhere
near the amount of resources to handle a large enough "neural net" for Civ. And if computers could
do that, you would not like the size of your electricity bill!

Human players can easily "think" many moves ahead in many scenarios and any AI won't be able to
beat them with PCs.

Here's just one very simple example to illustrate what I mean. (There are trillions of others.)
Suppose you have a scout in North America and you spot an opponent's galley headed along the coast
towards your civ in South America. You don't want that civ to find a City State and compete with
you. You know that you can easily buy a coastal tile and stop the galley from moving passed your city
in > 10 turns.

Do you think any AI (now, or for a very long time) is going to be capable of planning that far ahead,
anticipate what you are doing, or might do, and incorporate that into its tactics and overall strategy?
Chess is a far, far simpler game than Civ. For AIs they are not even in the same universe of possibilities.

(I've been using "AI" to optimize ships, wings and for other fluid dynamic applications for > 35 years.
Most (if not all) neural nets and AI algorithms implement some form of internal "optimization".)

Chess neural-net-based-AI can self-play thousands of games per day at a superhuman level on an average modern PC.
I agree that Civ is more complex than chess, but this is also true for human players (I am not saying that such an AI would play perfectly).

 

[Leyrann]

You don't want an AI that plays perfectly anyway. You just want an AI that tends to make good choices, if possible on a level comparable to the top 1% of players (assuming you have a method to make it artificially play worse for the lower difficulty levels).

Also, keep in mind you don't need to do the neural network learning on the computers of the customers. You do the neural network learning on a special computer, and then implement the results in the game.

 

[HorseshoeHermit]

An AI that could play Civ well would be a little bit discomfitting, considering the abstraction that it makes out of government edicts and war.

I think an AI is possible for the game, and the ideal disposition of it would be to play a game that creates fun (back to what Soren said). We're not talking about a general problem solver, the machine only has to have and utilize biases in the context of Civ's decisions, with a headstart given by some "rote" pathways and the tuning of the designers.