Slashman wrote:
The problem is exactly how to make a strong AI. A strong AI isn't an AI that's hopped up on insane bonuses so it can sidestep playing the game. It is an AI that is good both tactically and strategically that is difficult to achieve. If you can reach that point, then difficulty levels are actually meaningful. Because additional bonuses to the AI will now mean it will have more resources to do smart things instead of more resources to throw bad armies at the player in impossible numbers.



No developer refuses to makes the strongest AI they can. You can always gimp the AI by giving it less resources or earnings.



For instance. Giving the AI an extremely large/powerful starting army relative to the player and then telling it exactly where the player is and making it hostile to the player would likely result in more player defeats, but that's not strong AI.



So while you are right in one sense that the goal is to make the game fun, the AI acting like it is somewhat competent is a huge part of the immersion that can only really come from having a good AI in the first place. Because a bad one will be easily exploited and eventually disdained for providing no meaningful challenge.

I think you exactly miss the point. In a game, an AI will always be weak against a human intelligence, at least for several years still. So the point is exactly what myrec said : make an AI fun to play, which does not cheat too obviously and which play or try to play with the same rules as the player.

I agree, it would be a cool exercise And Caotico09 is right, it can be a way to handle small subsets of the AI problem. For the parts where it works, it can lead to nice effects, like unpredictable AI, like AIs that adapt to the meta but pleasantly lag smart humans... One of the things I would try to do, if I was in Amplitude's place, would be to gather as much data from real games as possible. Particularly multiplayer games, where you're already guaranteed a network connection.



I've always thought that one interesting way to dumb down AI would be to make it as smart as possible, but then lie to it about the game state. Add a random +/- to estimations of army strength, to travel times, to industry costs. Given a good enough AI to begin with, I think this would give a good simulation of a human that wasn't paying enough attention to the game. And if you ever test your AI by playing them against themselves on the server overnight, it would be a good way to introduce enough fudge factor to prevent narrowly tuned strategies from dominating and keep the AI feeling more human, less HAL.



AI is fun to think about.

natev wrote:


I've always thought that one interesting way to dumb down AI would be to make it as smart as possible, but then lie to it about the game state. Add a random +/- to estimations of army strength, to travel times, to industry costs. Given a good enough AI to begin with, I think this would give a good simulation of a human that wasn't paying enough attention to the game.

Oooh, I like the idea of adding in that margin of error, because when someone is playing they don't always exactly know/estimate what the other players are doing unless they are playing very close attention!

Regarding the "custom difficulty" direction, here is what I was used to do to enjoy a good and challenging game of EL (before Shadows) :



- Serious diffuclty for AI opponents

- Default settings for the world

- Pangea map

- Quest victory only



I know that some will say "blabla ... if you must deactivate some of the victory path, it means that the AI is broken and ... blablabla" but honestly the game and challenge feels very very good with this custom settings.



For Shadows I think I will try a new custom rule : I will try to not give order in battle and spectate it only, to be fair and play the same rules as my AI opponents (who also have the pain to have very bad tactical AI agents ^^).

Kruos wrote:




I know that some will say "blabla ... if you must deactivate some of the victory path, it means that the AI is broken and ... blablabla" but honestly the game and challenge feels very very good with this custom settings.



For Shadows I think I will try a new custom rule : I will try to not give order in battle and spectate it only, to be fair and play the same rules as my AI opponents (who also have the pain to have very bad tactical AI agents ^^).

Self-gimping FTW!

Slashman wrote:
Self-gimping FTW!

Very common in the 4x world.

DEFINITELY agree.



I like to play EL sparingly, once every 6 or so months, and EVERY TIME I run in to this issue. I havn't min-maxed a start yet so hard AI just completely overwhelms me, I can stay even on points with the normal AI and stay near the top of the pack, but then I feel dumb cause any war becomes so easy to win.



Need a new difficulty between normal and hard.

Problem is not to make Strong AI, problem is to make AI fun to play against. It's hard work and I think amplitude do their best.

myrec wrote:
Problem is not to make Strong AI, problem is to make AI fun to play against. It's hard work and I think amplitude do their best.

The problem is exactly how to make a strong AI. A strong AI isn't an AI that's hopped up on insane bonuses so it can sidestep playing the game. It is an AI that is good both tactically and strategically that is difficult to achieve. If you can reach that point, then difficulty levels are actually meaningful. Because additional bonuses to the AI will now mean it will have more resources to do smart things instead of more resources to throw bad armies at the player in impossible numbers.



No developer refuses to makes the strongest AI they can. You can always gimp the AI by giving it less resources or earnings.



For instance. Giving the AI an extremely large/powerful starting army relative to the player and then telling it exactly where the player is and making it hostile to the player would likely result in more player defeats, but that's not strong AI.



So while you are right in one sense that the goal is to make the game fun, the AI acting like it is somewhat competent is a huge part of the immersion that can only really come from having a good AI in the first place. Because a bad one will be easily exploited and eventually disdained for providing no meaningful challenge.

natev wrote:
The hard part of that is figuring out why the player is behaving in a certain way. If a player settles a particular tile, is it because of that tile? Because of nearby tiles? Because that tile is or isn't on a coast? Because there's a threat from an enemy unit? Because it's turn 20 and that tile is within reach of its settler? 4Xs are full of situations that will probably only ever happen once in the entire life of the release, across its entire audience. There are megabytes of data available to the AI at any given moment, only a few k of which are relevant to any particular decision.

That sums it up very well. Humans might follow an easy-to-see logical path to make a decision, but they might also make it out of whimsy or based on what they predict may happen 10 or 15 turns later. That's hard for an AI to assess. How do they know if that was a good or bad move for the player?

Kruos wrote:
I think you exactly miss the point. In a game, an AI will always be weak against a human intelligence, at least for several years still. So the point is exactly what myrec said : make an AI fun to play, which does not cheat too obviously and which play or try to play with the same rules as the player.

No. The point is that you can't do that. Do you know any modern 4x games where the AI cheats in something other than heuristic bonuses where it appears to be very smart and makes the player feel like they are playing by the same rules?



For the same reason that you state the AI will always be weak to a human intelligence, is the same reason that a cheating AI will be found out by the player and then exploited.



How do you make a dumb AI fun in diplomacy via cheats? How do you make a dumb AI make better strategic decisions via cheats? How do you make a dumb AI design better units via cheats? How do you make a dumb AI that doesn't know its own strengths or weakness better via cheats? You can't do those things convincingly without improving the AI in the first place.



It's the same reason that games like GalCiv or Sword of the Stars don't have weird cheats where the AI can do or know things the player can't. They end up making for a poor game experience once the game is past a certain level of complexity because if they are just cheats, they will break down once the player understands whats being cheated and how.

Slashman wrote:
No. The point is that you can't do that. Do you know any modern 4x games where the AI cheats in something other than heuristic bonuses where it appears to be very smart and makes the player feel like they are playing by the same rules?



For the same reason that you state the AI will always be weak to a human intelligence, is the same reason that a cheating AI will be found out by the player and then exploited.



How do you make a dumb AI fun in diplomacy via cheats? How do you make a dumb AI make better strategic decisions via cheats? How do you make a dumb AI design better units via cheats? How do you make a dumb AI that doesn't know its own strengths or weakness better via cheats? You can't do those things convincingly without improving the AI in the first place.



It's the same reason that games like GalCiv or Sword of the Stars don't have weird cheats where the AI can do or know things the player can't. They end up making for a poor game experience once the game is past a certain level of complexity because if they are just cheats, they will break down once the player understands whats being cheated and how.

It is exactly what we are saying here. So no need to play with words or '♥♥♥♥ the flies' (french expression - not sure of the translation ^^) around this topic, I see no added value in your post actually (no offense), everything has already been said here many times.

Kruos wrote:
It is exactly what we are saying here. So no need to play with words or '♥♥♥♥ the flies' (french expression - not sure of the translation ^^) around this topic, I see no added value in your post actually (no offense), everything has already been said here many times.

First of all. I initially replied to myrec's statement which said that the problem is not to make a strong AI but a fun AI. You really can't do one without the other IMO.



You then said I missed the point because AI will always lose to a human (which really isn't even absolutely true, because that depends on the competence of the human and their understanding of the game). I replied and asked how you can make a fun AI which is weak at the same time? Because as far as I can see, you cannot. Which is why I asked for examples of games that did this. No one supplied any.



My point was always that you can't achieve the fun part (in terms of making the AI fun to play against) without a good AI first. So how is that the same as what myrec said? Maybe it's a language thing but as far as I understand, the two statements aren't the same.

Not sure what to say guys I'm afraid- hence my hesitation to reply. We're aware that a number of people are unsatisfied with where the AI is at currently and we're making ongoing efforts to improve things. None of us can speak for the 4X genre as a whole though. I know there are a lot of very smart people out there working on the problem but I'm afraid that, lame as it may sound, it's a very hard one to solve.



If you have any specific problems with EL's AI please say so on the feedback thread. We do read this stuff! It's good to have it all in one place though.

Wilbefast,



We're not really criticizing you guys specifically here (at least not in the last few posts). More like having a philosophical discussion on what constitutes a good AI.



I have no doubt that you guys are working hard on it.



We all look forward to your efforts.

Slashman wrote:
First of all. I initially replied to myrec's statement which said that the problem is not to make a strong AI but a fun AI. You really can't do one without the other IMO.



You then said I missed the point because AI will always lose to a human (which really isn't even absolutely true, because that depends on the competence of the human and their understanding of the game). I replied and asked how you can make a fun AI which is weak at the same time? Because as far as I can see, you cannot. Which is why I asked for examples of games that did this. No one supplied any.



My point was always that you can't achieve the fun part (in terms of making the AI fun to play against) without a good AI first. So how is that the same as what myrec said? Maybe it's a language thing but as far as I understand, the two statements aren't the same.

Ok... I will replay.



First of all there is no hope for normal developers to create AI. AI is something different. Game developers are creating simple behavior patterns which can imitate human player. Problems arise when you want to make it decide. Many times you as a player decide according to "your feeling" computer behavior pattern don't have feelings. So it have to decide most of the time randomly or according to some predefined system. Even Amplitude posted a video about "Ai" in games. Look for it on youtube. You will realize that make good solid "AI" is not hard, but make it replayable and make it fun to play like 100 times is hard.



E.g. starcraft 1 had great "AI".. I mean strong one. They hard coded few strategies for each race and "ai" opponent execute that flawlessly, it's micro was better than koreans, it's macro was better than them too. So why you have never heard of that... cause after 30 or so games you realize what is "AI" doing and you outplayed it by abusing it's weaknesses. And that's game over. You will never feel satisfaction of game against that kind of "ai".



In 4X games it's harder to make strong "AI", cause there is no benefit for huge memory and compute power it have over the human player. So they have to invent "almost random" behavior patterns which will play the best. If you realize that even you as a player you play same "tactic" multiple times until you improve on that. "AI" will not have that change to improve over course of games.



How would you then make easy, normal, hard ? Number of strategies it have ? New players will never be able to beat even one of them. Trust me developers already have several strong strategies for each faction which are really valid and strong. Only you as a player discovered them and you can outplay them.



PS: look on youtube for videos about "AI" in games.

myrec wrote:




How would you then make easy, normal, hard ? Number of strategies it have ? New players will never be able to beat even one of them. Trust me developers already have several strong strategies for each faction which are really valid and strong. Only you as a player discovered them and you can outplay them.



PS: look on youtube for videos about "AI" in games.

In terms of what you want to call the behavioral patterns of the computer factions, whether AI or logic-based behavior patterns is semantics. Hopefully everyone understands that actual artificial intelligence is far away from us still. "AI" is easier to write and say.



In terms of difficulty levels, I have no issues with what has become the standard in 4x games currently. You make the best possible AI/behavior patterns you can and then use heuristic bonuses to scale levels up or down so that the best 'AI' behavior with no bonuses becomes the baseline or normal. Some developers gimp their AI by blocking the best algorithms to achieve lower difficulty. I think that's what FE:LH devs do in addition to lowering base income/mana/growth. But in general, to get higher difficulty, the norm is bonuses.



The basic problem that people have perceived with EL is that the baseline behavior for computer factions was still very weak even for a 4x AI. So stacking bonuses on it did not yield enough of an advantage to the computer factions. There are a number of reasons for that besides just 'humans are too smart'. Several have been outlined in the Feedback on AI thread. Some have gotten addressed with the Shadows patch and more are being worked on.

I think slashman is just trying to point out that Amplitude has the capability to come up with a better balance between fun and strong AI, closer to Civ than to Total War, and I agree. Sid´s had over 20 years working on that AI, though, and have rewritten it from scratch more than once. Time plays a large part.

Propbuddha wrote:
I won't defend the AI and agree that EL is complex, but not every game can be Chess. I think that depth and narrative/sandbox experience is what most people want from 4Xs (i.e. "Fun", not "Good").

Maybe that is so. I equate a fun 4x as something that is at least challenging and interesting. If not from strong AI, then at least the environment should offer a challenge.



If I'm looking for a good narrative, I can go play an RPG which will more than fulfill that need. In terms of sandbox, I may even find some RPGs which have that open world/do what you want feel like Mount and Blade. If I play a 4x, though, eventually I'm butting heads against the other players (AI). Sooner rather than later, I'm going to find out how empty that is if the other players don't know how to play. Narrative feels really flimsy if I'm unopposed in discovering it. Maybe if this game had some really dangerous native denizens like Fallen Enchantress: Legendary Heroes, I could tolerate the poor AI better. However, chances are that those denizens would simply eat the AI alive in short order.



So yes, not every game can be Chess, but we have moved so far away from chess level that we're essentially moving toward a point where it will be impossible to make any AI for the 4x genre if this continues. Pretty much every 4x game in recent memory has launched with an extremely poor or downright non-functional AI. Some of those got fixed to a point, others are still struggling months or years later. I'm not looking forward to having to wait a couple years after every game launch to comfortably play.

Have you tried normal AI difficulty with hard world difficulty?