I think that is true for the most part, but there are a lot of human characteristics you probably don't want in an AI like rage quitting, pointless attacks just to anger someone, feeding another civilization to spite someone, etc. IMO an ideal AI should behave similar to human intelligence, but it should not seek to entirely mimic it.
-
We are currently performing site maintenance, parts of civfanatics are currently offline, but will come back online in the coming days. For more updates please see here.
The difference between a System that is geared to fit the AI...
- Thread starter Ryika
- Start date
I think that is true for the most part, but there are a lot of human characteristics you probably don't want in an AI like rage quitting, pointless attacks just to anger someone, feeding another civilization to spite someone, etc. IMO an ideal AI should behave similar to human intelligence, but it should not seek to entirely mimic it.
Ryika
Lazy Wannabe Artista
- Joined
- Aug 30, 2013
- Messages
- 9,393
I actually really dislike the idea of human-like AIs. To me that's like trying to force it into a role that it can never fill, while at the same time giving up the strengths of what it can be - opponents that are part of the experience, that help immerse the player into the story that he writes for himself.
HorseshoeHermit
20% accurate as usual, Morty
I believe there's a disconnect between the Civ series and single-player AI design.
If we just look at "Some single-player 'builder' challenge with AI rivals", then there are lots of points that can be made. Having a complex field of choice, having AIs that can make good choices a lot of the time, but for which that field's nuances, situational variance, and mental-contest potential, will of course escape the AI, creates a game where the Human's ability to look at what is happening and think about it inventively will win him the day. That's a good game. That's a pass.
But civ games don't have that freedom to be such a game, to afford the conditions to an AI. The economics and tile exploitation and strategies that span a hundred turns to flower, there isn't an AI that plays competently short of one that plays amazingly. It just doesn't make sense, with what a strategy game is, a builder game is. The inescapable conditions of a 4X (e.g. snowballing - and the toolbox "negative feedback" is orthogonal to the snowballing that occurs in a builder game).
And civ games have a motive to be something different on the AI side - the historical leader facet. Or, just the general playing characteristically facet, in place of minmaxxing. This is Beyond Earth, but the 'characterizing behaviour' thing still exists, so you've got this voice to make them specifically not good players, but flavourful ones. (The trouble, see, is how sacrificing 'good' does not run it off the rails into abysmal whole cloth.) Now, a personality can be said to be a kind of default-state strategy, but nonetheless the AI is not a participant in the game the same way a Human will be. They'll have their personality and/or their strategy and that's that. They will always be a scribble at the bottom of a Civ game's potential for exploratory play.
Take for a moment that I am too quick to be sure such an AI is impossible - a good one. Let's suppose it could work. There would be some restrictions, and you would commit to making the game and AI together - I think that's the dialogue you were inviting in the OP? It's troublesome because these conditions in many places conflict with that which makes the game exciting were we to discount the PvAI mode:
Opportunity for bold advances stimulates creativity in exploiting the game, and also allows for the potential of multiple competing routes to success. Big numbers and swingy decisions, the audacity of multiple available slingshots keeps things in balance and depth for at least a long tail while the metagame is built up after release, ideally it in fact stabilizes a competitive and deep game in perpetuity. But meanwhile, slingshots are not friendly to an AI, which wants a flattened progression diagram, one that grows in requirements at a smothering rate, one that 'branches' rapidly and, whether it allows or heavily rewards long plunges or not, it must require breadth or the AI just will be left behind. If the AI is tinkered with to use the slingshots, then the developers would have to know the slingshots, which (A) defeats the excitingness of the bold strokes in the first place, and (B) defeats the excitingness of 'finding' them over just watching a computer use them, and (C) is improbable given well-known balance testing limitations.
Also the tactical combat. This is a laborious point, I'm sorry. Tactical AIs exist and are pretty darn good. But I claim it is the tactical system that screws up Civ's AI. 1upt makes the combat tactical, but 1upt is not the problem. The combat should be strategic. Or tactical. But at present it is trying to be both, and making a subsystem of the AI be good at a potpourri is ludicrous. Strategic combat would be about using the right terrain and fighting at the right time for the right geography (to protect or obtain territorial assets). The tactical combat is trying to use positioning, sight, mobility, surprise and all that to play a tactical game, which would be fine except you don't get to choose to not have a strategic combat game there, because it is there, and what those tactical games have in common where the AI challenge works is that everything else stays still and the grand strategy is outside the frame. a 4X AI has to be good at fighting war, and also good at figuring out how good a war might be, and also good at figuring out how much to want a war to fiddle with other things dozens of turns before that war would actually happen. That's just the basic economic reasoning of the genre.
As soon as you wash over that much complexity with "We'll let the AI spam resources", that's hands in the air and given up on what Ryika is asking. you need an (artificial reasoning) tool that , might not be good, but at least -tries- to do those things, and somehow, you need a game where an imbecile can make a decent go at those decisions, and yet you want to excite your player to tackle the same strategic difficulty to 'beat' it, meaning those decisions matter.
You somehow need to have a game where arbitrary decisions won't work, where random decisions are 95% confident not to work, (i.e., choices matter) and yet a rule-following machine can keep up with that same game.
So like I said. A single player , AI-driven challenge is totally doable. You can have it. It can be tough and fun. But it basically pulls in 180 degree opposite direction from making a Civ game you'd actually want to have.
The Civ 5 and Civ:BE did not just have an AI that "kept up" with some handicaps to shove it to keep tie with you. Those AIs are given literally everything, so that they don't play the game. They don't make choices. Not merely 'middling good' ones - they don't make any, they scoop up everything and are kept at bay by an artificially pacifist diplobalance.
As your experiment discovered, Ryika, once you take away some bold advance potential - flattening out the progression by making technology so much further away, and greatly expanding the requirements for advancement to include nearly every dimension - the AI got really good, and pure material won hard.
So the work cut out for even experimenting with meeting the AIs to the game's progression design, tying this back to your actual request for input, is to let the AIs make choices: Get them to where they are choosing things, and the Human is choosing things, and the AI isn't that good at seeing the future, or interpreting you, or managing gambles, but it is at least putting one foot in front of the other. Then you get to play along, do smart stuff in a string and beat the AI down and be proud of yourself. The AI gets cheaty to stand back up , and then you outsmart it again, and you do that throughout the ages and finally you prove your ultimate superiority. Good game. At present they are doing nothing strategic, so the task of tricking them into acting that way, combined with swapping the game around, is a tall order indeed.
If we just look at "Some single-player 'builder' challenge with AI rivals", then there are lots of points that can be made. Having a complex field of choice, having AIs that can make good choices a lot of the time, but for which that field's nuances, situational variance, and mental-contest potential, will of course escape the AI, creates a game where the Human's ability to look at what is happening and think about it inventively will win him the day. That's a good game. That's a pass.
But civ games don't have that freedom to be such a game, to afford the conditions to an AI. The economics and tile exploitation and strategies that span a hundred turns to flower, there isn't an AI that plays competently short of one that plays amazingly. It just doesn't make sense, with what a strategy game is, a builder game is. The inescapable conditions of a 4X (e.g. snowballing - and the toolbox "negative feedback" is orthogonal to the snowballing that occurs in a builder game).
And civ games have a motive to be something different on the AI side - the historical leader facet. Or, just the general playing characteristically facet, in place of minmaxxing. This is Beyond Earth, but the 'characterizing behaviour' thing still exists, so you've got this voice to make them specifically not good players, but flavourful ones. (The trouble, see, is how sacrificing 'good' does not run it off the rails into abysmal whole cloth.) Now, a personality can be said to be a kind of default-state strategy, but nonetheless the AI is not a participant in the game the same way a Human will be. They'll have their personality and/or their strategy and that's that. They will always be a scribble at the bottom of a Civ game's potential for exploratory play.
Take for a moment that I am too quick to be sure such an AI is impossible - a good one. Let's suppose it could work. There would be some restrictions, and you would commit to making the game and AI together - I think that's the dialogue you were inviting in the OP? It's troublesome because these conditions in many places conflict with that which makes the game exciting were we to discount the PvAI mode:
Opportunity for bold advances stimulates creativity in exploiting the game, and also allows for the potential of multiple competing routes to success. Big numbers and swingy decisions, the audacity of multiple available slingshots keeps things in balance and depth for at least a long tail while the metagame is built up after release, ideally it in fact stabilizes a competitive and deep game in perpetuity. But meanwhile, slingshots are not friendly to an AI, which wants a flattened progression diagram, one that grows in requirements at a smothering rate, one that 'branches' rapidly and, whether it allows or heavily rewards long plunges or not, it must require breadth or the AI just will be left behind. If the AI is tinkered with to use the slingshots, then the developers would have to know the slingshots, which (A) defeats the excitingness of the bold strokes in the first place, and (B) defeats the excitingness of 'finding' them over just watching a computer use them, and (C) is improbable given well-known balance testing limitations.
Also the tactical combat. This is a laborious point, I'm sorry. Tactical AIs exist and are pretty darn good. But I claim it is the tactical system that screws up Civ's AI. 1upt makes the combat tactical, but 1upt is not the problem. The combat should be strategic. Or tactical. But at present it is trying to be both, and making a subsystem of the AI be good at a potpourri is ludicrous. Strategic combat would be about using the right terrain and fighting at the right time for the right geography (to protect or obtain territorial assets). The tactical combat is trying to use positioning, sight, mobility, surprise and all that to play a tactical game, which would be fine except you don't get to choose to not have a strategic combat game there, because it is there, and what those tactical games have in common where the AI challenge works is that everything else stays still and the grand strategy is outside the frame. a 4X AI has to be good at fighting war, and also good at figuring out how good a war might be, and also good at figuring out how much to want a war to fiddle with other things dozens of turns before that war would actually happen. That's just the basic economic reasoning of the genre.
As soon as you wash over that much complexity with "We'll let the AI spam resources", that's hands in the air and given up on what Ryika is asking. you need an (artificial reasoning) tool that , might not be good, but at least -tries- to do those things, and somehow, you need a game where an imbecile can make a decent go at those decisions, and yet you want to excite your player to tackle the same strategic difficulty to 'beat' it, meaning those decisions matter.
You somehow need to have a game where arbitrary decisions won't work, where random decisions are 95% confident not to work, (i.e., choices matter) and yet a rule-following machine can keep up with that same game.
So like I said. A single player , AI-driven challenge is totally doable. You can have it. It can be tough and fun. But it basically pulls in 180 degree opposite direction from making a Civ game you'd actually want to have.
The Civ 5 and Civ:BE did not just have an AI that "kept up" with some handicaps to shove it to keep tie with you. Those AIs are given literally everything, so that they don't play the game. They don't make choices. Not merely 'middling good' ones - they don't make any, they scoop up everything and are kept at bay by an artificially pacifist diplobalance.
As your experiment discovered, Ryika, once you take away some bold advance potential - flattening out the progression by making technology so much further away, and greatly expanding the requirements for advancement to include nearly every dimension - the AI got really good, and pure material won hard.
So the work cut out for even experimenting with meeting the AIs to the game's progression design, tying this back to your actual request for input, is to let the AIs make choices: Get them to where they are choosing things, and the Human is choosing things, and the AI isn't that good at seeing the future, or interpreting you, or managing gambles, but it is at least putting one foot in front of the other. Then you get to play along, do smart stuff in a string and beat the AI down and be proud of yourself. The AI gets cheaty to stand back up , and then you outsmart it again, and you do that throughout the ages and finally you prove your ultimate superiority. Good game. At present they are doing nothing strategic, so the task of tricking them into acting that way, combined with swapping the game around, is a tall order indeed.
Making smarter AI for CiVBE is relatively simple and straightforward, however that requires a lot of effort and programmer time (e.g. resources that I suspect Firaxis doesn't have).
About a year ago they had an opening for "AI programmer / Scripter" and job name alone says how little Firaxis is devoted towards creating good AI.
Its simple to write code that will find good position for city based on some rules (determining such rules is a little bit trickier) and will decide what impovements to build / switch citizens between tiles to maximize output (just like human do) and etc.
Making good battle AI is harder. But nothing too hard.
My guess is that Firaxis simply lack enough resources.
Source: my experience. has been coding game AI =)
About a year ago they had an opening for "AI programmer / Scripter" and job name alone says how little Firaxis is devoted towards creating good AI.
Its simple to write code that will find good position for city based on some rules (determining such rules is a little bit trickier) and will decide what impovements to build / switch citizens between tiles to maximize output (just like human do) and etc.
Making good battle AI is harder. But nothing too hard.
My guess is that Firaxis simply lack enough resources.
Source: my experience. has been coding game AI =)
HorseshoeHermit
20% accurate as usual, Morty
That's some good sleuthing, but maybe their project structure involves some senior architects who design and coders who implement. Like, for instance, the project structure that Firaxis is known to have. Programmers don't need to have vision, they need to be technically literate. Division of labour amen.
Anyway, our subject is what to even aim for, what an AI experience can be, which places Firaxis' actual situation on quite a tangent. I merely observed that the AI to work from is in the pitiful position of not even approaching intelligence or any differentiated behaviour, which hobbles a "mod" approach to bringing a game to one's goals. Either the AI kinda plays the game you are , with handicaps, or the AI controls an environment which purposefully is not the same as the player's situation. Players moan about "I SEE YOUR ARMY ON MY BORDERS" but it belies the generality.
I don't like talking about inconveniences I want to talk about goals. There is some overlap with a concrete case of actual AIs in advancing a point about ideal ones, that's all.
Hail
Satan's minion
that would be a great game! with an AI playing to loose.
the civ5 AI's inability to actually play the game, wrecks the experience for me far more than 1UpT.
matrix multiplication is relatively simple and straightforward, but will require a lot of effort to get right.
writing an AI is hard. it is a very complicated task to debug and to force the AI to avoid doing stupid things.
it is trivial to assign weights to yields and then pick the best tiles based whatever tile has the max dot product value. however, all this math has nothing to do with an AI. it is a dot product.
an AI must plan, memorize it's doings and learn on it's mistakes.
no. for the classical monetizing model (*) investing time and effort into an AI is not the optimal way to spend the, rather limited, budget. there is simply not enogh bang for the buck.
(*) the classical model:
- release game N (for civ5 N=5).
- release some DLC.
- release an expansion.
- repeat steps 2 and 3 several times.
- ???
- PROFIT!
- N = N + 1.
- goto step 1.
I predict that some software may move toward the software as a service approch, since M$ chose to incorporate it in Windows 10 (win 10 will be the last windows. also that is why M$ allows people to upgrade from win 7/8 to win 10 for free for the first year.).
with the software as a service approach, things change. the basic idea is to have one game/application that is continously updated. to play/use a game, a user must buy a subscription (for a month, a year, etc.). a robust AI will prolong a game's replayability, keeping the players' interest.
an AI should, obviously, know how to play the game. the better the AI, the more SP and MP converge, which may be undesired, if one does not want to balance the game.
I believe, that as the difficulty level raises, the AI should play more optimally.
Galgus
Emperor
- Joined
- Aug 22, 2012
- Messages
- 1,705
Good point, the AI should be a respectable opponent.
In Civ games, they also shouldn't totally break immersion.
Unless its Gandhi, apparently.
Ryika
Lazy Wannabe Artista
- Joined
- Aug 30, 2013
- Messages
- 9,393
Well, that's not much different from what MMOs already do.
And seeing what is going on in WoW, with the developers basically doing the baseline-work to keep enough people playing so they can buy the precious store-mounts and other paid services, I'm very critical towards the idea of having that kind of model become the "standard" in other genres.
However, I think that old approach won't work much longer either, now with steam refunds, an Arkham Knight that has been ported so badly that the Publisher decided to pull the game after apparently getting so many refunds that it hurt their baseline. Especially with the huge amount of players being very cynical towards BE - I don't think a Civ 6 can allow itself to be "mediocre at best" if they want players to keep the copies they had pre-ordered.
Galgus
Emperor
- Joined
- Aug 22, 2012
- Messages
- 1,705
The willingness of people to buy new Civ games is based on the trust and goodwill from previous, good releases.
If they continually release half-baked games that need two expansions to be fleshed out, it will damage that trust.
Personally BE's release has me skeptical on RT beyond what we absolutely know will be included.
_______________________
That said its important to differentiate between a half-baked game that needs two expansions and a good game that just gets better with more expansions.
If they continually release half-baked games that need two expansions to be fleshed out, it will damage that trust.
Personally BE's release has me skeptical on RT beyond what we absolutely know will be included.
_______________________
That said its important to differentiate between a half-baked game that needs two expansions and a good game that just gets better with more expansions.
Hail
Satan's minion
making games is a business.
the TBS and RPG genre can leech more money from players using the subscription model (with DLC and expansions, obviously!). a decent AI is required for such TBS game. rule bending for the AI should be avoided at all cost. it's vastly immmersion breaking.
games with lower replayability, notably, like the infamous The Order: 1886, will likely stick to the classic model.
HorseshoeHermit
20% accurate as usual, Morty
The presentation you linked
Though the things he says on the 'good AI' side appear to reduce to "We don't actually know how to play strategy games, and we're surprised when our algorithms make the AI dumb", it is good work.
Yet no 'developer resources' influx is gonna help with the competence problem, unless they meant hiring in someone(s) to do AGI.
I said simple. Not easy.
You can copy paste solution from the book or use any 3-rd party free library. How is this a lot of effort?
Sure matrix multiplication is time-consuming but as far as I know better algorithm doesn't exist.
Using weights for such task is an extremely bad idea.
Been there
Practice shows that when you have to choose the best location for something based on many rules you can't use weights for an adequate choice.
I remember I have seen the AI-controlled soldier taking obviously bad position during the firefight, ignoring much better position 2 meters away. Why? -becaue the better position had weight 10, while his current position had weight 1 000 000 for some mysterious reason
The weights coefficients were tweaked perfectly for one scene, and totally didn't worked for the other (there were 20+ different weight modifiers; the game like civilization would probably require much more - it would be a nightmare).
On the contrary, well-thought data representation, decomposition of abstract-looking tasks and smart algorithms can create an illusion of the "true" AI good enough.
"True" AI has nothing in common with video-game AI.
That's what behaviour trees are used for.
patches
You can learn a lot from the replays of a good players.
If I didn't like the game I would not buy its DLC's.
E.g. Endless Legend had horrible interface and was quite boring - so I wouldn't buy its DLC - although it had some interesting ideas so I probably would buy Endless Legend 2 if they ever make it.
Personally I think that the best monetizing model is Fallout 3 aka "Collect them all" - plenty of small DLC's that expand an original and big add-on.
For civilization BE it could be new biomes, new leaders, etc.
But I agree with you that usually developers and publishers are not interesting in creating a good AI since it is expensive and invisible - it can't be shown in gameplay trailers (I really hope that early-access mode can change that - generate some hype long before release - okay that was wishful thinking).
You don't need something intelligent to do what is considered an intelligent work. Think car navigator. Or trip planning-booking-car renting. As I said, well-thought data representation, decomposition of abstract-looking tasks and smart algorithms can create an illusion of the "true" AI good enough.
Heck I remember most of the human players played Settlers 7 (it was multiplayer-focused - one of the best multiplayer game I ever played in fact - StarCraft of 4X =) ) worse than AI's of that game, and they followed the same rules as a human player.
Ryika, love your mods. My Idea is to increase their cost and spacing requirements of academies/nodes etc. as you go to higher difficulties. Also on the lowest levels, where you would face the current AI, you would face Sue Fielding. At higher levels you would face Suzanne Fielding, which would be with your Awesome improvement. At the mid levels you would face Suzanne, but she would not have the strategic resources that she "liked".
Similar threads
