Better AI in CIV 6 ?

[Alphons Rodulfo]

I would not mind having a slider or some other setting allowing me to sacrifice turn times for better AI. At least for some parts of the AI there must be coded numerical limits.

 

[Magil]

I disagree with many of the comments here on the nature of AI. I think people are confusing the concepts of an "equal" with an "obstacle."

I don't think anyone has any particular expectations of the AI being "equal". In general, you don't see many complaints about how the AI operates at the strategic level, despite them generally being bad at it and only being carried by bonuses. I mean, for one thing, it's just plain harder to notice.

It's when the AI embarrasses itself in combat that people really notice. 1UPT has been the AI's anathema. Notice that mods like Improved AI and/or Vox Populi tend to make the AI better at tactical combat, but it's still not on the level of the human player--and I think few people really expect it to be. They just want it to not make stupid moves, they want it to be able to move and shoot, to retreat wounded units (and perhaps attempt to cover the retreat using zone-of-control), and to not embark units right next to a bunch of ranged units. They want it to prioritize threats to its empire, not throw everything it has at a city-state or barbarian camp while the capital falls.

 

[Growth]

This is further complicated by the fact that in order to be fun, 4X AI has to play irrationally. There is nothing rational about agendas; they inherently make AI unequal. Agendas are actually weaknesses for exploiting. Played rationally, the AIs would ditch their agendas and be aware they were in a game.

I don't see it from this perspective. See, there is nothing rational about wheat or fish being near my capital. I see agendas like that, it's part of the circumstances and context. So assuming agendas are part of the context, the real question from the AI perspective should be:

Given that:

• I got embodied as e.g Montezuma.
• I got this and that UA, UU, UI, agenda, etc.
• I spawn next to a river.
• I am next to Scythia, France and Brazil is the human player.
• There is Crab and some milk here.

How can I optimize my situation and environment and which victory type best fit my context?

You can clearly be rational about all of that, barring Combinatorial Explosions, and as a human you should too. Now, that doesn't mean it couldn't be affected by some flavors like mood, time of day and moon phase to make them more unpredictable.

What bothers me with Bonus aided AI among other things, as I am sure others can relate is that sometimes they can achieve something, for instance build the Temple of Artemis, and not only could you not have been able to build it in your situation, you couldn't possibly have built it being given their situation in same turn time (minus the bonuses).

I can accept being beaten at a confrontation or at a race to a world event based on dice rolls but not because of a bonus, that's just cheap. Many players enjoy being outsmarted by an AI at a game, it's enlightening and makes you reflect on your own shortcomings, none of that great stuff happens when you find out they've been cheating. I think the performance of the AI isn't that much relevant in many games, but something like Civilization, it's something else entirely, the stakes are high and there is a huge import on the gameplay, as a matter of fact, some people won't even buy the game based on how they predict the AI won't be able to cut the mustard.

 

[qwerty25]

I already said this in another thread: If the AI can move and shoot in one turn, ill consider it a success.

Sent from my LG-H850 using Tapatalk

 

[Ryika]

the real question from the AI perspective should be:

Civ 5 already did that to an extend with Flavors, everything more than that would be questionable at best, because stuff that cannot be done (properly) with flavors inevitably requires custom code for every situation. That's a lot of work for little effect - and more importantly makes the AI very dependent on the standard set of rules, which would be HORRIBLE news for modding.

What bothers me with Bonus aided AI among other things, as I am sure others can relate is that sometimes they can achieve something, for instance build the Temple of Artemis, and not only could you not have been able to build it in your situation, you couldn't possibly have built it being given their situation in same turn time (minus the bonuses).

That's partly a result of Civ 5s frontloaded AI bonuses and partly an inevitably. The AI will never be able to adapt at the same rate that a human player can, so it must be able to do things that you cannot do in return, otherwise it has no chance to even give the illusion of being a tough opponent. However, if the AI didn't start with multiple technologies and huge %-boosts, instead gained more and more boosts over time players would not feel as if they have no chance at getting anything during the early game (on high difficulties).

 

[Aheadatime]

However, if the AI didn't start with multiple technologies and huge %-boosts, instead gained more and more boosts over time players would not feel as if they have no chance at getting anything during the early game (on high difficulties).

That's why I made a mod for myself which increases the deity AI's % bonuses (unit train time, building train time, growth, etc), but removes their free settler and techs at turn 0. Fun stuff so far.

On topic, I don't think the AI will be dramatically improved in Civ6. I'm quite certain the AI will move and shoot finally, but that's about it. The pre-release footage I've seen confirms this for me. Tons of wasted movement, bad strategic choices, lack of terrain manipulation, etc.

Combat aside, I think these agendas are going to cause diplomacy to be a sh#* storm tbh. I can't fathom getting the stars to align in such a way that there is a pack of 2 or 3 game-long allies anymore. Everything you do will upset somebody... build wonders, improve your luxes, settle land (lol), declare war on your forward settling neighbor, have tons of culture, have no culture, etc. This applies to each AI as well. I foresee tons of denouncing, surprise wars, temporary alliances resulting in backstabbing as modifiers fluctuate, etc. Going to be a mess imo.

 

[Thormodr]

I think the AI for Civ VI will be better than Civilization 5. Mind you, it is pretty hard to get any worse.

Now, I never said that the AI in Civ VI will actually be good but poor is still better than abysmal. Lol.

 

[darko82]

I think the AI for Civ VI will be better than Civilization 5. Mind you, it is pretty hard to get any worse.

Now, I never said that the AI in Civ VI will actually be good but poor is still better than abysmal. Lol.

As I said, nothing to be proud of as a creator or artist. This is not an achievement to be proud of. That said.

People at Firaxis should make the AI to be proud of themselves.

But it seems they do not care or they are not capable of this achievement.

 

[Thormodr]

As I said, nothing to be proud of as a creator or artist. This is not an achievement to be proud of. That said.

People at Firaxis should make the AI to be proud of themselves.

But it seems they do not care or they are not capable of this achievement.

I do hope that they'll take particular pride in the AI. Time will tell.

 

[ashvin.l]

Hmmm that article reeks of sensationalism

If you read into it it mostly makes word association. Kicking ass is click baiting.

I actually read the original paper. The new AI wins the default free-civ AI 78% of the times. If that's not kicking ass, I dunno what is.

 

[ashvin.l]

Machine learning is a cool possibility--ultimately kind of similar to how something like Google functions. But also vulnerable to problems of its own. For one thing, once the AI discovers something undesirable it can actually be difficult to make it unlearn that thing since all of its strategies are connected. For example, if the AI learns that one particular Research path always produces the most favorable results. That may be effective, but its also boring to players. It may also reveal game imbalances, but let's be real here--game balance is never going to be perfect.

To say nothing of what happens when players learn that "AI always chases Wonder X." Well, Wonder X might be ideal in a bee line sense, but is it still best if opponents anticipate you will chase it every time? Is it still best when there are 8 of you all following the same strategy, because it is "the best," but only one of you can grab the wonder? This is not straightforward question. The AI navigating a maze, and the walls of the maze are changing.

Then the next part: what happens when the game is modded? Unless the AI can learn on the fly, it's strategies will no longer work.

I think it's much better to think of AI as the "level" you are up against. Montezuma is "the Aztec-themed level" for example. Difficulty settings change some things about how it plays. The level has some exploits coded in for the fun factor (getting mad about Luxuries for irrational reasons, unconnected to the outcome of the tantrum). You just aren't ever going to get an equal player out of an AI in this sort of game, not with today's technology anyway.

Now if we just mean AI at the tactical/minor strategy/ unit level, you can see some of that in the Better AI mod, which is pretty reasonable, and about as good as it's going to get.

I would disagree with the assessment. Machine learning is not one-size fit all solution. Especially in Civ 6, where what you do depends heavily on the map. The AI using machine learning algorithm would probably solve something like a greedy optimization problem (which is to enhance the immediate score or get an advantage, etc). The answer to that problem would be highly specific to where the game is, what are the Civs strengths, what is the map, etc. If the AI is learning from users and users DONT do the same thing over and over again, then I would find it unlikely that the AI would do the same thing, independent of the situation.

Secondly, while modding is an important aspect of the game, Firaxis is responsible for the AI they build for THEIR game NOT for the modded game.

Lastly, I would love to see the AI identify game imbalances. If there is one choice which AI always follows no matter what the situation is, then it represents a serious imbalance issue and must be corrected. No body said we can't use AI to fix bugs

 

[SMcM]

Most players case little about the AI, so it usually has low budget. Civ 6 AI will, undoubtedly, be underwhelming to those who hope for a good AI.

Also, is that a joke-poll? Allowing multiple answers and not having an answer for a better AI?

Yeah, right on both counts. It is unlikely there will be huge improvements for AI, but at least give us a proper poll!

 

[Ulthwithian]

It's not appropriate case for machine learning. You can't make "play to win" AI with condition and map size like this. So the AI has completely different approach with completely different goals like "don't make moves which look stupid". You just can't make this work with machine learning. Not to mention constantly changing rules during development.

Yes, that would be some of the 'other issues' I mentioned.

Frankly, the biggest--even insurmountable--issue is the question of an appropriate training set. Who do you tune your AI for? And, unlike some views, this isn't simplistic in the least.

 

[stealth_nsk]

Yes, that would be some of the 'other issues' I mentioned.

Frankly, the biggest--even insurmountable--issue is the question of an appropriate training set. Who do you tune your AI for? And, unlike some views, this isn't simplistic in the least.

That's why you don't need machine learning here. Simpler algorithms will fit better, especially those where you could adjust most of the behavior with data, not touching actual code.

Building order, city location and city specialization could be done with simple prioritized list algorithm. It's not very complex and allows throwing in a lot of modifiers without making unmanageable mess.

Strategic planning could be done with decision tree or similar algorithms.

Tactical AI is the most tricky one. I thought about the right approach and it looks like you can't avoid some bruteforce algorithms here. To make it calculatable, it needs to check only limited area of a single "front"; calculate 1 turn only, but add strategic positions not only pure unit state (similarly to how chess programs work). In this case, however, moving and syncing fronts becomes quite complex task.

P.S. After I thought about this, I realized one quite simple thing. If Civ5 tactical AI was done in the same way, its inability to move and shoot was quite simple error - it just wasn't included in the list of possible actions. The only reason why it wasn't fixed later I could think of is - the vanilla AI was fully written by Jon Shafer. So it's totally possible developers who came after him just didn't want to dive into those areas of his code.

 

[Ulthwithian]

As I said, nothing to be proud of as a creator or artist. This is not an achievement to be proud of. That said.

People at Firaxis should make the AI to be proud of themselves.

But it seems they do not care or they are not capable of this achievement.

Frankly, darko, it seems that you want Firaxis to make an AI that you can be proud of.

Also, to your earlier comment, there is a difference between making a 'good AI' and making a 'challenging AI'. You want the game to be challenging to you. Well... are you asking for personalized AI? That's not likely to happen.

You said that you wanted the AI to 'be able to take a city'. Well, okay, that's important... for a Civ that expects to win militarily. If the strategy selection doesn't create that 'win condition' or whatever for that Civ, should it even try to take cities?

I get your concern, but I don't think you realize the issues this spawns.

I am not a very good Civ player. I, frankly, would not purchase the game if I thought that every time I played it I would be in for a grueling 'live or die' situation.

At the end of the day, I highly doubt the condition of the CivVI AI on release is going to be a question of 'can Firaxis make a 'good'--for some subjective definition of good-AI, but whether they want to'. You take as given that they should want to, and I'm not seeing a lot of insight that your requirements for CivVI are just that: requirements. And that everyone has theirs, too.

I am pretty confident that no one can write an AI that can challenge both you and me. (Basic game theory tells you that an optimal strategy against optimal play is NOT optimal against suboptimal play.) Given that from Firaxis' standpoint, we are equal, who do they challenge? The typical way around this is through 'bonuses'. 'Bonuses' are not AI, however, and unless you write an AI to use the bonuses effectively, you can actually have a situation where the AI written to optimize Prince play actually does WORSE in Deity play.

So let's try to quantify requirements. I am going to assume, since you want an AI that can challenge you or someone you would identify as 'the typical Civ player', that you would be comfortable with using game win percentages as a metric for the AI strength.

So my question to anyone who has concerns about the AI in Civ:

If you are playing Single-Player, with N other Civilizations in the game, what do you believe the expected win % should be for the player in order to call the AI in the game 'challenging'?

 

[Acken]

The challenge is created through bonuses. If you dont want a live or die game play on a lower difficulty. If you want a hard game with the possibility of losibg play at a higher one.

But the base AI has to be good enough for the bonuses to provide a challenge in an interesting way.
If the AI is total crap that it needs more than double your yields and triple your units this is challenging but stupid.

The civ5 ai also uses some small tricks for the ai to have a different behavior on top of bonuses. Nothing big though.

 

[Ulthwithian]

That's why you don't need machine learning here. Simpler algorithms will fit better, especially those where you could adjust most of the behavior with data, not touching actual code.

And you can't parameterize machine learning algorithms with data? I certainly wouldn't use any algorithm that cannot at least have a data-manipulable front end by (say) modders.

Building order, city location and city specialization could be done with simple prioritized list algorithm. It's not very complex and allows throwing in a lot of modifiers without making unmanageable mess.

Sure, until someone learns the algorithm. This leads to 'AI always rushes Wonder X, so avoid Wonder X'. The issue with prioritized lists is that it's a static element that can be 'gamed against'.

Now, something that may help with Civ VI is the decompression of cities and Eurekas. Almost everyone in the Eureka thread seems concerned with the 'oh no are we going to have to get all Eurekas, or are they meaningless, etc.?' issue, and few people are talking about how this may help to get better AI.

E.g., I could use a simple Greedy algorithm--almost not even worth the name--that does Tech and Civic selection based on the 'cheapest' / 'least time to target' approach. In CiV, this would lead to a breadth-first tech acquisition strategy, which is apparently not very challenging for the advanced human player.

However, in CiVI with Eurekas, a Military-focused AI (e.g.) would get Military Eurekas more often, and that would make the Greedy algorithm naturally prioritize Military techs.

Strategic planning could be done with decision tree or similar algorithms.

Sure. Those don't seem to help much in CiV but as noted above, districts may change this. Without thinking too hard on this, if a City can manage to build the buildings that go into any one district (including City Center) in the time it takes to gain the population for the next district, then build orders (for buildings at least) are remarkably straightforward.

Tactical AI is the most tricky one. I thought about the right approach and it looks like you can't avoid some bruteforce algorithms here. To make it calculatable, it needs to check only limited area of a single "front"; calculate 1 turn only, but add strategic positions not only pure unit state (similarly to how chess programs work). In this case, however, moving and syncing fronts becomes quite complex task.

Enh... ironically, tactical AI is probably far easier to write than strategic AI, given that there is a wealth of material where tactical AI exists but not strategic.

P.S. After I thought about this, I realized one quite simple thing. If Civ5 tactical AI was done in the same way, its inability to move and shoot was quite simple error - it just wasn't included in the list of possible actions. The only reason why it wasn't fixed later I could think of is - the vanilla AI was fully written by Jon Shafer. So it's totally possible developers who came after him just didn't want to dive into those areas of his code.

Could vanilla AI be modded? If so, that means that it can't be that hardcoded. In fact, the most valid criticism of CiV AI is that modders were able to make it challenging using the AI modding API.

 

[Ulthwithian]

The challenge is created through bonuses. If you dont want a live or die game play on a lower difficulty. If you want a hard game with the possibility of losibg play at a higher one.

But the base AI has to be good enough for the bonuses to provide a challenge in an interesting way.
If the AI is total crap that it needs more than double your yields and triple your units this is challenging but stupid.

Are you aware that the optimal AI to efficiently use one set of resources may not be the optimal AI to efficiently use a different set of resources, even if one set is a subset of the other?

I would ask that people answer the question at the bottom of my post, though. What do you believe your win percentage should be in an n-player game of Civ that has a challenging AI?

 

[stealth_nsk]

Sure, until someone learns the algorithm. This leads to 'AI always rushes Wonder X, so avoid Wonder X'. The issue with prioritized lists is that it's a static element that can be 'gamed against'.

It depends on the number of parameters. Civilization favor, city specialization, available tiles and so on, so forth. Even without random flavors it's nearly impossible to predict the AI behavior in such system.

Now, something that may help with Civ VI is the decompression of cities and Eurekas.

As I wrote in eureka thread - the AU eurekas are nearly invisible to human player, so playing them fairly with AI is just not needed.

Enh... ironically, tactical AI is probably far easier to write than strategic AI, given that there is a wealth of material where tactical AI exists but not strategic.

I'm familiar with Chess AI principles. I could say you just can't calculate move on the whole field of Civ map. Which brings tactical AI to be way more difficult. And, unfortunately, that's the field where AI behavior is the most visible to player. AI left archer exposed to melee attack? It's called dumb and immersion is immediately broken. So no, it's not easy.

Could vanilla AI be modded? If so, that means that it can't be that hardcoded. In fact, the most valid criticism of CiV AI is that modders were able to make it challenging using the AI modding API.

Yes, I understand this. It's the question of amount of work and how well it adopts to changing rules. Also, you can't ignore the factor of developer interest. Diving into some other developer's code to fix some nasty bug is not what most of developers like to do. It's possible the AI developers gave very high estimation for fixing this bug, so its priority was pushed backwards.

But these are wild guesses, of course.

 

[Acken]

Are you aware that the optimal AI to efficiently use one set of resources may not be the optimal AI to efficiently use a different set of resources, even if one set is a subset of the other?

I would ask that people answer the question at the bottom of my post, though. What do you believe your win percentage should be in an n-player game of Civ that has a challenging AI?

Im aware of a lot of stuff tyvm.

My statement still holds as I never asked for optimal behavior on all levels. The ai algorithm in civ5 still scales pretty well with its bonuses so its not even really the question. Its weakest point being diplomacy and tactical combat. Which neither is affected by bonuses.

Your question is meaningless. It asks a random statics varying from every person which in turn will vary on the difficulty they play at. At best id say i shouldbt win 100% on the highest level.