3982173917524862 wrote:
I would like to ask if the improved AI will be exclusive to ES2 or will it be also improved in ES1 and EL?

It is very unlikely that an AI system would be ported to other games. It just isn't that simple. :c

JackAu wrote:
I think AI can be achieved via cloud assisted system that rips off player play style historical data very much like that seen in Driverta thing in Froza games in xbox one, packets of some human players play style and settings be downloaded to player systems as special game AI tweak for that set game, so each new game would be different

While I understand the principles behind it in a racing game with set maps etc where you just have to lift somebody's trajectory, a 4x has a LOT of moving parts, on top of a randomly generated map in ES's case. I think a system aping another player's behavior would be extremely complex to design and implement.

I think that the most important thing is that the AI be not integrated and dispersed within the game. There must not be AI logic mixed within the game mechanics and configurations. This is a flaw I see too much in strategy games... This makes that any change in the game mechanics ends up breaking the AI. The AI should behave like a player, and have the same informations that a player have. What a player sees through the user interface, an AI player should be able to see through an API... And then takes decisions over that.

This means that the AI itself should have their own persistence data (that is, the AI must have "memory" ). Also, the AI should be able to interpret data and reach a decision, independently of AI hints: that is , the AI should decide by itself if a unit is good to attack, or to defend, for example... There must not be *hints* embedded in the game universe information, telling that this unit is for attack, and that unit is for defense... This is up to the AI to decide.



I have already suggested this for Endless Legend, but too late in the development cycle... Let there be an interface for plugging in new, third party, AIs, and you will end up giving an opportunity for a lot of college people who is actually researching and/or learning about AI to implement possibly wonderful new behaviors for new, wicked, devious AI personalities...

abmpicoli wrote:
There must not be AI logic mixed within the game mechanics and configurations. This is a flaw I see too much in strategy games...

Now, I'm no AI expert, but you say this is a general flaw in too many strategy games. So as an exercise, could you mention a few examples of games - preferably with the scope of a 4X - that uses this kind of methodology?

From my understanding, the methodology we used in Endless Legend is similar to that of most other prominent games of the same genre, such as Total War and Civ. So also as a general request to you all: What games of the same scope/genre has an AI you consider good? What parts of that AI is giving you the impression it is working well?

The problem with this topic ... is that I'm switching categories in ArXiv or google scholar... from superconductivity(Thesis...) to ... Artificial intelligence...



There some litteratures for everyone: https://www.cs.auckland.ac.nz/research/gameai/publications/Robertson_Watson_AIMag14.pdf



And why academic work is not implemented: http://phdcomics.com/comics.php?f=1815

Metalynx wrote:
Now, I'm no AI expert, but you say this is a general flaw in too many strategy games. So as an exercise, could you mention a few examples of games - preferably with the scope of a 4X - that uses this kind of methodology?

From my understanding, the methodology we used in Endless Legend is similar to that of most other prominent games of the same genre, such as Total War and Civ. So also as a general request to you all: What games of the same scope/genre has an AI you consider good? What parts of that AI is giving you the impression it is working well?

Off the top of my head:



The Sword of the Stars series. The devs have always strongly believed that the AI should not know or do things that the player cannot. SotS 1 is where I would look. SotS2 got really crappy treatment from the publisher and released too early.



Fallen Enchantress: Legendary Heroes. The AI does not cheat apart from bonuses past a certain level. And I might add that the AI on that game without bonuses is leaps ahead of EL. Well except for the Dead. They need some real work, but most of the others are more than decent.



Age of Wonders 3. Someone will no doubt tell me it isn't a 4x, but it pretty much has all the boxes ticked. It is just focused more on military power. Also a very decent AI that offers a challenge without cheating. If you go past Knight difficulty, the AI does begin to get bonuses to production, gold and mana but nothing else.

Metalynx wrote:
So also as a general request to you all: What games of the same scope/genre has an AI you consider good? What parts of that AI is giving you the impression it is working well?

Galactic Civilizations 2, no doubt. Not only do factions behave differently, but in all 'personalities' the AI is capable of engaging in diplomacy in ways that are as of yet unrivalled.



The AI is not that great when it comes to war however, and is not optimal when it comes to economy and building infrastructure properly. But in terms of diplomacy, it's the best one i've experienced, though naturally a human player can still exploit the AI diplomatically.

Jiheu wrote:
The problem with this topic ... is that I'm switching categories in ArXiv or google scholar... from superconductivity(Thesis...) to ... Artificial intelligence...



There some litteratures for everyone: https://www.cs.auckland.ac.nz/research/gameai/publications/Robertson_Watson_AIMag14.pdf



And why academic work is not implemented: http://phdcomics.com/comics.php?f=1815

This isn't the only reason that academic work isn't implemented - games are often late too



I remember seeing, back when I was a student, the Bos War learning AI and wondering why more games don't use Machine Learning and other "advanced" AI techniques. The answer is that ML means a black-box: an AI that is incomprehensible and highly unpredictable. Our job as game designers is to create an interesting experience for our audience - this goal is not always in line with the goals of academia. Academics and game designers have very different answers to the question "if an AI falls in a forest and nobody hears about it, is it stupid?".



It's really important to get external perspectives though. We are in touch on and off with a couple of experts both from industry and academia: their view-points are highly appreciated And players see things from yet another angle, which is why your input is so useful (EARTH)!

I think that making a good AI in a 4x is not necessarily creating a single AI that can match a human player(this is pretty much impossible anyway in anything as complex as a 4x game), but rather that the collection of AIs together should provide an interesting and challenging game together with the game's environment.



Cheating is definitely not desirable (the AI should not know or do things that the player cannot). But even more important, is that the AI should not NEED huge bonuses to play a decent basic game. This is currently not the case in either ES or EL and it is the main turn off for me and why I don't play either game anymore.

Slashman wrote:
I think that making a good AI in a 4x is not necessarily creating a single AI that can match a human player(this is pretty much impossible anyway in anything as complex as a 4x game), but rather that the collection of AIs together should provide an interesting and challenging game together with the game's environment.

The feeling at Amplitude is that, while having some sort of "Director AI" could help dynamically balance the difficulty and prevent uninteresting situations (like the two AIs crippling eachother in a three-player game), it just wouldn't be "kosher" for a 4X game to have what is essentially an "AI Game Master". In other words 4X games are more about simulation than narrative - they are about emergence from a system of strict rules. The focus is not on procedurally generating cinematic experiences based on fuzzy logic. But what do you folks think?

Slashman wrote:
Cheating is definitely not desirable (the AI should not know or do things that the player cannot). But even more important, is that the AI should not NEED huge bonuses to play a decent basic game. This is currently not the case in either ES or EL and it is the main turn off for me and why I don't play either game anymore.

Metalynx looked into a few of the 4X games most often quoted as have "AI that doesn't/rarely cheats" - turns out that these AIs are generally just better at hiding the fact that they're cheating They'll do things like cheating for information about what you're building and then building the counter to that. Players don't notice these things, they just think the AI is being smart, and these sorts of cheats make a huge difference to the power of the AI despite said AI not actually being any smarter. In EL the initial goal was to not cheat at all if possible, so the AI can't see things it shouldn't be able to. This may have been overly ambitious though, and ultimately the resource bonuses needed to be added in to give AIs little extra help, especially on higher difficulties.

wilbefast wrote:
The feeling at Amplitude is that, while having some sort of "Director AI" could help dynamically balance the difficulty and prevent uninteresting situations (like the two AIs crippling eachother in a three-player game), it just wouldn't be "kosher" for a 4X game to have what is essentially an "AI Game Master". In other words 4X games are more about simulation than narrative - they are about emergence from a system of strict rules. The focus is not on procedurally generating cinematic experiences based on fuzzy logic. But what do you folks think?

wilbefast wrote:


Metalynx looked into a few of the 4X games most often quoted as have "AI that doesn't/rarely cheats" - turns out that these AIs are generally just better at hiding the fact that they're cheating They'll do things like cheating for information about what you're building and then building the counter to that. Players don't notice these things, they just think the AI is being smart, and these sorts of cheats make a huge difference to the power of the AI despite said AI not actually being any smarter. In EL the initial goal was to not cheat at all if possible, so the AI can't see things it shouldn't be able to. This may have been overly ambitious though, and ultimately the resource bonuses needed to be added in to give AIs little extra help, especially on higher difficulties.

I think it's an interesting discussion, knowing that in a game the players will have access to meta-information the AI wouldn't necessarily know if it "didn't cheat", such as their previous experiences, their knowledge of the behavior of such and such builds or research paths, the presence of such and such other species in the game, etc. I'm not sure how an AI can emulate this or if we even would want them to. Do I want an AI to adapt its expansionist tactics to the size of the maps so it rushes out earlier on smaller maps?



Those are fascinating questions, and I'd like to be a fly sitting on the wall in these AI meetings to see what you guys come up with to keep us hooked.

Yeah I think a (possibly optinal) game master AI is something that does deserve attention. The issue is often within victory conditions and larger goals that necessite some specific and previsible scripts. I've heard people complaining it was really difficult to keep allies when approaching victory since you can't have an allied victory per se.



More generally, writing an AI that knows how to play the game is something difficult while the game is in progress. But it still something mandatory sine most player experience will be vs the AI. Letting the door open for custom AI scripting could be a nice move, since there are people eager to tune this.



Another thing iI miss often is offering customisable difficulty settings. Rather than have the whole package : AI more agressive, with more resources, more techs, and less hapiness for you, you could at least pick separated "AI nastiness" and "GAME rules nastiness" parameters.

wilbefast wrote:
The feeling at Amplitude is that, while having some sort of "Director AI" could help dynamically balance the difficulty and prevent uninteresting situations (like the two AIs crippling eachother in a three-player game), it just wouldn't be "kosher" for a 4X game to have what is essentially an "AI Game Master". In other words 4X games are more about simulation than narrative - they are about emergence from a system of strict rules. The focus is not on procedurally generating cinematic experiences based on fuzzy logic. But what do you folks think?

I'm not sure I want an AI game master either. I don't think my post was suggesting that. At least it wasn't my intention. I prefer that interesting gameplay emerges because the AIs do what they need to do (pursue their objectives) and the world itself also moves along helping or hindering based on chance.



If people prefer a more directed experience, then I would suggest the inclusion of scenarios. This would allow the devs to create situations and world setups that make things harder or more interesting for the player with a specific narrative goal in mind. I don't think I wan't that on randomly created maps.

Metalynx looked into a few of the 4X games most often quoted as have "AI that doesn't/rarely cheats" - turns out that these AIs are generally just better at hiding the fact that they're cheating They'll do things like cheating for information about what you're building and then building the counter to that. Players don't notice these things, they just think the AI is being smart, and these sorts of cheats make a huge difference to the power of the AI despite said AI not actually being any smarter. In EL the initial goal was to not cheat at all if possible, so the AI can't see things it shouldn't be able to. This may have been overly ambitious though, and ultimately the resource bonuses needed to be added in to give AIs little extra help, especially on higher difficulties.

Can you please specify which games these were? I'd like to know which on my list falls into these categories. I have personally spoken to the devs from Age of Wonders 3 who have said that they do not make their AI cheat. Also the ones from Sword of the Stars. So if you have found that this isn't true, I'd like to bring it up with them and see what they say.

Slashman wrote:
Can you please specify which games these were? I'd like to know which on my list falls into these categories. I have personally spoken to the devs from Age of Wonders 3 who have said that they do not make their AI cheat. Also the ones from Sword of the Stars. So if you have found that this isn't true, I'd like to bring it up with them and see what they say.

Firstly I should point out: Heuristic based systems like we use is inherently cheating. Though selectively so (based on implementation). For example, no 'fog of war' exists, but we sometimes selectively ignore units. When you calculate an 'empire threat' level, you will often times use the military stats of the player to find that value - regardless of vision and such. But when you make military movements, you use vision - for example.



Sword of the Stars

http://wiki.swordofthestars.com/sots1/AI_--_Modding - This is the same type of AI we use. Heuristic based where you set values for different things in the game.

http://wiki.swordofthestars.com/sots1/Game_Setup#Difficulty_settings - From normal and up the AI gets 1.5 to 3.4 times faster research and up to 1.7 times economic bonus.



FE:LH

http://forums.elementalgame.com/450250 <- this link implies that it uses the same heuristic based system we use.

http://steamcommunity.com/app/228260/discussions/0/828939163776914379/ - The opponents get mainly 'combat' bonuses. More health and attack stats.

Also found mentions of better starting conditions and some small economic bonuses - but no clear sources.



AoW

This game is amazing difficult to find information on. Think the best link I found is: https://steamcommunity.com/app/226840/discussions/0/611703999970647865/?insideModal=1

Essentially: They gimp the AI (make it take stupid decisions) below what would be 'normal'.

Above 'normal' they give economic bonuses. Exactly what, I cannot say. But something in between 'flat economic bonuses' as in EL to 'starting with more cities' and getting 'better combat stats'.



EDIT: Found this:

As for the bonuses, from lord on, every level gives the AI 100 Base Happiness, 33% income, and 2 healing per turn for undead/machines. King and Emperor get +2 Def/Res against independents.

EDIT: To be clear, income in this case is Gold, Mana, Knowledge and Production, while population is excluded.

Source: https://steamcommunity.com/app/226840/discussions/0/535152276585909661/



Gal Civ 2 (which was mentioned - don't remember by who)

http://galciv.wikia.com/wiki/Difficulty_level - Essentially same as AoW.

Low difficulties - Intentionally making bad decisions for AI.

Above normal: flat economic bonuses 105% all the way to 400% on extreme difficulty modes. Also the AI cheat by selecting ship designs that directly counter the players ^^





Essentially it is the same for most games. The main thing I found out, is that the games people complain the least about try to focus on one thing:

Give more 'starting bonuses' instead of 'constant high bonuses' - and hope that the AI's start will snowball them ahead, without making them impossible to catch up to.



Also I found out, that if you search AI on any 'grand strategy/4X' game - you will find more people complaining than praising. But I guess that is the same for everything in this world! Though it appears that AI 'personality' is a large factor of praise.

Another thing iI miss often is offering customisable difficulty settings. Rather than have the whole package : AI more agressive, with more resources, more techs, and less hapiness for you, you could at least pick separated "AI nastiness" and "GAME rules nastiness" parameters.

If we can tweak an IA, it will be good. Could be an option to balance the IA between three major parameters Economy/ Research/ Army (something like that). The AI game master could be interesting (can he force the differents AI to be allied versus the human player[ifheistoopowerfull])?). Most of the time, the problem, with AI, is the diplomacy. They switch their attitudes every time with no apparent reason (allied for 100 turns and suddenly become enemy because you are close to his frontier). Some stronger coalition after a number of turn could be good.

There was an interesting mod for Civ V I played where the mod designer had changed the way AI bonuses worked. Rather than it being a massive increase of resources at the beginning or being done by age. The extra resources where tied to how well the AI was doing compared to the player.



For example if the AI had more gold income than the player then they would get no bonuses. However if the player was getting more gold than the AI then the AI would get a gold bonus.



It was like this for the rest of the resources.



I enjoyed playing with this mod more than the usual Civ V difficulty settings of 'give the AI more stuff' because it made the game more interesting.



It felt more like a level playing field yet a 'rubber band mechanism' meant that when I was steam-rolling ahead the AI bonuses kept them in the game and kept the challenge up.



Was good fun.

Hey Metalynx,



In those games I listed, I did mention that from certain difficulty levels, they do get bonuses to things like production, gold and research. However, the difference between say AoW 3 and EL, is that I can find a decent challenge from the AI on the default difficulty level with no bonuses. Same for Sword of the Stars and the same for Fallen Enchantress: Legendary heroes.



In EL, I can literally crap around for numerous turns and still come out on top. It has happened every game I have tried after each patch which had AI improvements. One of my previous AoW 3 games actually had me tie with one of the AI for an Ascension win after some awesome back and forth battles and alliance switching and breaking etc. I don't get similar experiences in EL if I play with no bonuses. The same for Fallen Enchantress: Legendary Heroes.



One of the things that FE:LH does is have the AI with the ability use the equipment loadouts of human players from past games. I have played games where the AI uses my saved designs for a certain race from a previous game (they had the same name and everything). It's actually quite effective because the AI will sometimes have issues where 'designing' units is concerned. I don't know how feasible it would be to add this to EL or ES 2.

Ah I know Slashman The comments made me look into the examples put up for context. I'm definitely not trying to outright defend our AI - I was trying to understand why you (and possibly others) enjoyed those AI, by looking at technical implementation and numbers.



I generally think the issue you are describing with AI on normal difficulty in EL - is that it will very rarely win before the score timer runs out. I actually strongly believe that they make rather decent decisions (outside of battles) - they are just not working towards a goal - i.e. a victory condition. The higher difficulty ones get to some of the victory conditions (science/economic/diplomatic) due to the bonuses they receive - then the issue is that players cannot keep up at all during the game because of these bonuses.

But that is just my 'theory' for now ^^

A good A.I that is programmed and understands how to play the game.Not oh lets give it huge cheats on low levels ala ES1 on release.It was still easy to beat and destroyed the game theme,immersion.

Better than battlefield 4 bot....

Metalynx wrote:
Ah I know Slashman

I generally think the issue you are describing with AI on normal difficulty in EL - is that it will very rarely win before the score timer runs out. I actually strongly believe that they make rather decent decisions (outside of battles) - they are just not working towards a goal - i.e. a victory condition. The higher difficulty ones get to some of the victory conditions (science/economic/diplomatic) due to the bonuses they receive - then the issue is that players cannot keep up at all during the game because of these bonuses.

But that is just my 'theory' for now ^^

Honestly that may be a big part of it. In Age of Wonders 3, for instance, the AI will visibly pursue a Seal or Beacon victory when it becomes powerful/rich enough. That naturally invites conflict because AIs that were peaceful or even allies with that particular faction will definitely declare war and break alliances. This is especially true if there is no Allied Victory condition enabled. The other AIs all react accordingly. This then creates a very interesting flurry of activity as seal positions and cities are assaulted and allies change. Former enemies may team up to stop the player who is closest to winning. A major difference is that the tactical AI in AoW is really decent compared to the EL. It casts spells, dispels and disjuncts sensibly and it uses flanking and unit abilities well. Diplomacy is very simplistic in AoW, but it functions well.



What I find boring/bad is having to give so high a set of bonuses that the AI ends up playing a different game from the player, or more specifically, it doesn't have to play the same game the player is.