TotalBiscuit thinks the AI is good. and i'm now willing to believe it...

[Smith of Lie]

The problem is that the best algorithms we have are awful. There is no need to make them worse for anything but the very lowest difficulties (settler, warlord). Its easier and cheaper to give the human bonuses at those difficulties, and good enough.

What you are asking for is a better AI at high difficulties. We all want that. But saying it needs to be done is entirely moot as nobody knows how to do it in anything resembling sensible time and cost scales.

We could always aks IBM to teach Watson how to play Civilization and try to copy that. But that might be slightly resource intensive. And impossible to copy outside the lab. But certainly it would be interesting experiment.

 

[Olleus]

We could always aks IBM to teach Watson how to play Civilization and try to copy that. But that might be slightly resource intensive. And impossible to copy outside the lab. But certainly it would be interesting experiment.

Wouldn't work

As I recall, Watson uses a deep learning neural network. That requires absolutely vast data sets for it to learn how to do something. For something like recognising a picture of a cat or playing Go, we're talking in the hundreds of millions of different inputs. Considering all the different variables in Civ and that a good strategy (ie, rushing a religion) might be great in one game (say, starting as Russia in some tundra), but could be absolutely awful in another (starting as Monty surrounded by enemies). So you would need a lot of different inputs to distinguish them.

A incredibly rough back of the envelope calculation (will post if anyone is interested) suggests that if you had civ games running continuously on 1000 different cores in would take 50 years to have enough data to feed to Watson for it to be able to have ago. Probably not worth it.

 

[God of Kings]

Wouldn't work

As I recall, Watson uses a deep learning neural network. That requires absolutely vast data sets for it to learn how to do something. For something like recognising a picture of a cat or playing Go, we're talking in the hundreds of millions of different inputs. Considering all the different variables in Civ and that a good strategy (ie, rushing a religion) might be great in one game (say, starting as Russia in some tundra), but could be absolutely awful in another (starting as Monty surrounded by enemies). So you would need a lot of different inputs to distinguish them.

A incredibly rough back of the envelope calculation (will post if anyone is interested) suggests that if you had civ games running continuously on 1000 different cores in would take 50 years to have enough data to feed to Watson for it to be able to have ago. Probably not worth it.

It would be much easier for Watson to do Civ I instead.

 

[Txurce]

The problem is that the best algorithms we have are awful. There is no need to make them worse for anything but the very lowest difficulties (settler, warlord). Its easier and cheaper to give the human bonuses at those difficulties, and good enough.

What you are asking for is a better AI at high difficulties. We all want that. But saying it needs to be done is entirely moot as nobody knows how to do it in anything resembling sensible time and cost scales.

For zero cost, and over a significant amount of time, the Community patch (now VP) has a kickass AI that doesn't even require many of vanilla's bonuses. There's no comparison to BNW. Since Firaxis had both funds and time, we should expect a competitive AI at higher levels, even on the inauguarl version. And if they don't deliver it, it's because they chose not to... probably because there aren't enough high-end players.

 

[Cerilis]

For zero cost, and over a significant amount of time, the Community patch (now VP) has a kickass AI that doesn't even require many of vanilla's bonuses. There's no comparison to BNW. Since Firaxis had both funds and time, we should expect a competitive AI at higher levels, even on the inauguarl version. And if they don't deliver it, it's because they chose not to... probably because there aren't enough high-end players.

Think of it like that: Is the kickass AI also beatable for the average casual non-civfanatic that never plays higher than prince? ^^"

 

[Txurce]

Think of it like that: Is the kickass AI also beatable for the average casual non-civfanatic that never plays higher than prince? ^^"

On Chieftain, maybe. And maybe not.

I agree with your overall point. That's why I said that Firaxis may simply be choosing not to improve it as much as they could.

 

[bbbt]

And if they don't deliver it, it's because they chose not to... probably because there aren't enough high-end players.

This is definitely the case - if we go by the Civ 5 achievements, only about 15% of players play on a level higher than Prince.

 

[Acken]

You'd have to remove people that haven't played the game at all or barely of that statistics.

The truth about steam is that a lot of games are bought and never played.

 

[gunnergoz]

Maybe so but I'll bet Civ is not one of them...unless we are speaking of bargain bin buyers who just collect games. That I could see. But anyone that buys a $60 game at retail is going to be serious about playing it.

 

[Acken]

Just look at the stat of people not having the achievement for finding an ancient ruin if you don't believe it. And then it drops from there really quickly. Very simple achievement have relatively low global completion.

 

[agonistes]

I've only been considering the problem for about 3 minutes now, and this may be slightly off topic as my post isn't commenting on the quality of CVI's AI, but looking at the question of the feasibility or difficulty of preparing an AI to operate differently at different difficulty levels of the game...

I'm by no means a professional programmer. I've only done scripting for Neverwinter Nights, all self-taught, but I HAVE done AI scripting there. I just never considered working on an AI that changed in difficulty - I always predetermined how challenging I wanted a certain 'encounter' to be, and programmed the AI accordingly. Its not a great walk from NWN AI to Civ AI, I certainly can see how I would script out Civ's AI. What I find attractive about the problem is the AI scaling. How would I go about it? Could it easily be done? Its entirely possible that CVI has already done this, isn't it?

To me, a good AI for Civ would have some way to recognize and respond to human behavior. Obviously it couldn't account for every possible behavior/action (and this, of course, is one reason why a decent player should always win). And, of course, the AI would have pathing for achieving victory. And the two structures would have synergy. GCIII, and other games I imagine, grant AI civs bonus computing power at higher levels of difficulties, but that's not part of this discussion. I'm talking only about the actual coding.

One idea would be to start with a very terrifyingly capable AI script, and at certain points of 'decision' give it a % chance of following through with a decision. The chance would be close to 100% at deity. Then work backwards. A second would be to include in the higher level of difficulties possibilities for actions for the AI to take that it wouldn't at lower levels of difficulty. And yet a third would be to have the AI have preset actions to take at the higher levels of difficulties. Probably a good idea to mesh all of this with the various leader personalities and agendas, but the higher the difficulty, the more commonly seen some of the behaviors would have to be.

So, yes, fairly easy to do I think. Time consuming, but easy. Probably the biggest challenge would be getting AI programmers who are skilled enough at playing the game that they would be able to code an AI capable of dealing with highly skilled players. As we have seen in every version of Civ, one of the biggest advantages the player has is simply efficiency. Never has the Civ AI been capable of matching human efficiency with regards to resource management, be it hammers, commerce, or workers. Why not? The fact that a human can steamroll a deity Civ even with the deity bonuses proves that the imbalance is there, and massive. The AI should be coded in such a way that optimally the major advantage the human player has is that he or she is able to switch strategy and tactics on the fly - the ability to successfully change dynamically in response to situations - more quickly than the AI can. The only reason I can think of for this glaring discrepancy between human efficiency and AI efficiency is that the programmers are coding AI without knowing enough about how resources should be managed. Possibly not a comment on their playing abilities, as, after all, they most likely are far into the programming while much of the 'concrete' game is still being made.

I'm not too worried about the AI in CVI though. From the viewpoint of someone who has tinkered with AI, it appears that a number of the agendas and abilities were created with AI in mind. I expect we will see some major AI updates in the first few patches because the dev team WANTS to have a good AI, and have been paying attention to that throughout the design process. But you aren't going to have a challenging AI at deity right off the bat, not with such a limited amount of testing at that level at a point where the rest of the game was fairly polished. AI tweaking while the rest of the game is being worked on gets to be a real black hole... a push here becomes a pull there sort of thing.

As for some these AI decisions we are seeing in the videos, well, its Prince. PRINCE. Maybe something to worry about, maybe its not. Because there is the possibility that there is nothing to worry about, and the ramifications if there IS something to worry about I don't really see as WORTH worrying about, I'm not going to worry about it.

 

[Lord Shadow]

As it's been said earlier, scaling the AI to difficulty isn't feasible, as the "best" game-grade AI just isn't all that good. There's simply no case of an AI being "too good" purely by virtue of its ability that dumbing it down would make sense.

So it's always in the player's experience's best interest that the AI plays at its best.

 

[agonistes]

As it's been said earlier, scaling the AI to difficulty isn't feasible, as the "best" game-grade AI just isn't all that good. There's simply no case of an AI being "too good" purely by virtue of its ability that dumbing it down would make sense.

Its feasible in that the AI *could* be good enough. I can't explain *why* its not good enough. As I mentioned, there is the glaring discrepancy between human efficiency and AI efficiency which has persisted for quite some time now. Just patching that hole would make a huge difference. But they never have. Which leads me to believe that they either lack the ability as programmers, lack the ability as players, or don't have the time. I suggested towards the end of my post that it was a time issue. But scaling is most certainly feasible. If I can do it feasibly, any real programmer could.

Although I suppose if its a question of feasibility, then, by definition, a lack of time in and of itself would make it unfeasible. So let's just call it doable.

 

[Zhahz]

Well, I'd like the AI to not do boneheaded things like send out unguarded settlers, especially when it tempts players (or make it a city state/bigger diplo hit so that it isn't so trivial to grab setters/workers).

That press build had to be a few months old and hopefully some of the glaring issues have been cleaned up (things like city states flooding units too).

Overall though, AI is a tough thing and especially when trying to keep turn speeds within reason, but since I play pretty casually and don't always reduce the game to a diety war game I'm usually fine with the AI for the most part.

The devs played some on emperor this week and the AIs were settling at an alarming speed along with building lots of units and being aggressive. Some good discussion and happenings in that video, including them trying and failing an early rush.

 

[Lord Shadow]

Its feasible in that the AI *could* be good enough. I can't explain *why* its not good enough. As I mentioned, there is the glaring discrepancy between human efficiency and AI efficiency which has persisted for quite some time now. Just patching that hole would make a huge difference. But they never have. Which leads me to believe that they either lack the ability as programmers, lack the ability as players, or don't have the time. I suggested towards the end of my post that it was a time issue. But scaling is most certainly feasible. If I can do it feasibly, any real programmer could.

Although I suppose if its a question of feasibility, then, by definition, a lack of time in and of itself would make it unfeasible. So let's just call it doable.

It's a very big hole, and there's the added challenge that AI calculations have to be fast as well so as not to make turn processing times unbearable.

You're really underestimating the challenge AI faces in a very complex game such as Civ if you think you can remotely easily come up with a "very terrifyingly capable AI script", one so good (and fast enough) you'd have to scale it down several notches lest only Deity-level players could beat it.

 

[Txurce]

It's a very big hole, and there's the added challenge that AI calculations have to be fast as well so as not to make turn processing times unbearable.

You're really underestimating the challenge AI faces in a very complex game such as Civ if you think you can remotely easily come up with a "very terrifyingly capable AI script", one so good (and fast enough) you'd have to scale it down several notches lest only Deity-level players could beat it.

I don't want to beat this drum any more, but... the VP mod provides enough challenge for all players, at one level or another, and with fewer bonuses than vanilla.

 

[GhostSalsa]

As we have seen in every version of Civ, one of the biggest advantages the player has is simply efficiency. Never has the Civ AI been capable of matching human efficiency with regards to resource management, be it hammers, commerce, or workers. Why not? The fact that a human can steamroll a deity Civ even with the deity bonuses proves that the imbalance is there, and massive.

Deity players just leech off the AI bonuses, eg worker stealing and chain-bribes. Paying one gem to have two mindless giants destroy each other doesn't make you David. Which is not to say that human ingenuity can't shine in playing civ, just that it has nothing to do with the difficulty level. A player who plays sub-optimally on Emperor, where tech/gold/city siphoning tactics don't work, and gets swarmed by a runaway AI army, and loses cities, needs more creativity than anyone finishing a game at -260. I don't agree that the malfunction of high difficulty levels in civ has to do with human efficiency, it has to do with a game system that allows template-able optimal strategy at all, which leads to expecting the AI to act as a human-level opponent by way of a limiter, which it can't do.

The AI should be coded in such a way that optimally the major advantage the human player has is that he or she is able to switch strategy and tactics on the fly - the ability to successfully change dynamically in response to situations - more quickly than the AI can. The only reason I can think of for this glaring discrepancy between human efficiency and AI efficiency is that the programmers are coding AI without knowing enough about how resources should be managed.

We can't program AI's to change responses to situations before we can even program them to conceive situations as such. All the AI can be programmed to do is react to values. The bulk of civ's AI dev work is just refining how values are assigned to discrete player actions, and superimposing psuedo-strategic conditionals for how and whether the AI changes agendas based on those values, without ever, ever, ever, ever, being a thing that is understanding what the game is.

 

[agonistes]

This:

All the AI can be programmed to do is react to values.

tells me we can do this:

program AI's to change responses to situations

Which you seemed to agree with, when you said this:

refining how values are assigned to discrete player actions, and superimposing psuedo-strategic conditionals for how and whether the AI changes agendas based on those values

You can program the AI to be able to do anything except that which you don't program it to do.

But I apologize for disagreeing with the majority (which I wasn't aware I was doing), and, I fear, taking the thread off topic.

 

[nzcamel]

I believe you're the only one complaining about this. Too strong cities is generally viewed as one of the biggest problems in Civ5.

Maybe they've gone too far in the opposite direction given that in 1UPT you can only have one defender inside the city center; but I like their intent; which is to make it again risky to settle cities without defenders.

The devs are on holiday now, probably the game is getting on gold this time around.

Your tongue is in your cheek right!? I doubt the top Devs have moved far from their desks recently, nor will for another week or so.

 

[stealth_nsk]

You can program AI to do a lot of things (not anything, of course), but the more complex AI becomes, the more it's dependent on the exact game balance and values. That's significant. If developers release some balance patch, they want to be able to change only couple of values in the AI to be in sync with the patch. If they have to reprogram the AI for each patch (which would surely involves a lot of testing), this would hurt patching schedule a lot. Say, instead of monthly patches we'll get patches every 3-5 months (with the same amount of changes). Eventually this would lead to much less developed game in the end of its cycle.

It's all about compromises.