Does the AI plan ahead at all?

One thing I notice in multiplayer games is that good player plan ahead to a good degree. They have their faction's questline memorized so they know to make what they'll need to continue on in the quest before they need it. I play as the vaulters, and I know I always level up my first hero in such a way that shed be a good governor because the questline, at one point, demands a vaulter hero with a thrice leveled up governing skill.



Come to think of it.. how good is the AI about following its questline? Some really good boosts come from doing stuff it tells you to do...

HappyHead wrote:
Does the AI plan ahead at all?

The AI doesn't simulate X turns into the future, as this would require far too much computational power. Nor does it plan out what it's going to build or research many turns ahead, as this would be inflexible. It does however plan army manoeuvres: to send an army to besiege a city means outfitting the units, building them, assembling them, marching them to the destination, laying the siege and, eventually, storming the city walls. It also prepares things like empire plans and diplomatic contracts in advance: if it wants war or an alliance it will save up points and put them aside.

Propbuddha wrote:


Rapidly negotiates Peace in any war that AI didn't start

Interesting nuance here: the RC AI personality can't just give its peace ministry more votes, as this will make it less able to impose its will late game. Or will it? If the AI is smart enough to trade with everyone and really lines its pockets then maybe making it rain mercenaries (even if they're not a great return on the investment) could do the job.

Propbuddha wrote:


Preferes to fight battles in Forest areas (Forest Rage is a big boost)

Tricky to do, but that's an interesting point. I've taken note Of all the other suggestions to - I don't necessarily reply every day but I am reading this

As it has just now dawned on me that I am supposed to post AI feedback here, I am crossposting from this thread: /#/endless-legend/forum/5-general/thread/2500-drakken-ai-instigating-war

clykke wrote:
I am gonna grab this chance to offer some feedback.



I don't think it is inherently a good thing that each faction play a certain way, though it certainly adds immersion if done correctly. Actions of the AI should always be based on what would benefit it the most at any given time and play to it's general strengths, regardless of any "role" it may have in the lore. For instance, the Drakken does not really benefit in particular from peace. They do not have any affinity or traits that should make them want peace. On the contrary, the Drakken have strong units and good commanders, and actually have the advantage of playing the war game with lesser risk than anyone else. They even have a quest requiring them to gain control of an enemy city. They should be warmongers as soon as it looks like they have the upper hand, and fall back on their affinity if needed. From a cost (risk)/benefit perspective this makes a lot more sense than being peaceful builders/techers. There is no reason to only take advantage of the ability to force peace in a passive/defensive way.



The Drakken should basically play the game like a human would (meaning adapting as needed and playing to the strengths of the faction), and if that conflicts with their lore, it should be regarded a design flaw and possibly corrected.



Every faction specific AI should be designed this way, though it is okay to add a little bit of randomness in order to avoid predictable (and thus exploitable) behavior. Random personalities could also easily be added as an option.



Each faction should play to the strengths of the faction the same way a human would do, as long as it is not detrimental to do so (Nechrophages should never declare war on a (militarily) stronger enemy, unless that enemy is already at war). They should also understand that deviating from their standard play style for a short time can often improve their chances of success in whatever area they excel. For instance, Necrophages should understand that peaceful expansion improves their capacity to wage war at a later stage.



I am not sure exactly how your "ministries" work, but I did something similar with the AI for a small (and bad) RTS as the final project when studying. I believe the Necrophages should work something like this (obviously simplified):

1) Head of state: Set priorities: Wage war (standard behavior for the faction)

2) Head of state -> Military Ministry: Any viable target(s)? -> No

3) Head of state: Set priorities: Build units, Wage war.

4) Head of state -> Financial Ministry: Can we afford more units -> No

5) Head of state: Set priorities: Improve economy, Build units, Wage war.

6) Head of state -> Worker Ministry: Allocate workers to dust.

7) Head of state -> Financial Ministry: Can we afford more units -> No

8) Head of state -> Production ministry: Can we build improvements that improve our economy? -> Yes

9) Head of state -> Worker Ministry: Allocate workers to production.

10) Head of state -> Production ministry: Build economic buildings.

11) Head of state -> Financial Ministry: Can we afford more units when current construction is done -> No

12) Head of state -> Expansion Ministry: Any nearby regions to settle? -> Yes.

13) Head of state -> Production Ministry: Build a settler.

14) Head of state -> Expansion Ministry: Settle a region.

15) Production ministry -> Head of State: Order completed (Build settler)

16) Production ministry -> Head of State: Order completed (Economic buildings)

17) Expansion ministry -> Head of State: Order completed (Settle region)

18) Head of state -> Financial Ministry: Can we afford more units -> Yes

19) Head of state: Set priorities: (Improve economy removed): Build units, Wage war.

20) Head of state -> Production Ministry: Build units.

21) Production ministry -> Head of State: Order completed (Build units)

22) Head of state -> Military Ministry: Any viable target(s)? -> Yes

23) Head of state: Set priorities: (Build units removed): Wage war.

24) Head of state -> Military Ministry: Any viable target(s)? -> Yes

25) Head of state -> Military Ministry: Go to war.



Something like that would make the Necrophages a warmonger at heart, but make them smart about it. They will have different priorities at times, but only to accomplish what they really want: Conquer all of Auriga.

Every faction should have their own (not necessarily unique) "standard priority" based on their strength (not lore), which would make them fulfill a certain role, but at the same time allowing them to adapt to all kinds of situations. They will only pursue their favorite strategy when it is wise to do so, but will do it whenever possibly.



This is, in my opinion, the best way of implementing different AI personalities without making them predictable, exploitable or downright stupid.



Thank you for reading. It got a little longer than intended. Sorry about that...

When I've been attacked by AI while already embroiled in another war, I've thought, "If the AI did that on purpose, it was smart." If the AI does this on purpose, this would be a good place to put some flavor text so that players know that it wasn't just chance.

natev wrote:
When I've been attacked by AI while already embroiled in another war, I've thought, "If the AI did that on purpose, it was smart." If the AI does this on purpose, this would be a good place to put some flavor text so that players know that it wasn't just chance.

OOH, yes. Something to the effect of: "good luck fighting on two fronts, loser" [declareswar]

or "I'm sorry, were those yours?" [coldwar]

Just thinking a little more about it-- wilbefast, you mentioned that sometimes the AI seems dumb, and you look in its innards and it's actually being smart. It sounds like what you're saying is it'd be smart to add flavor text to make the player understand what the thought process is. Even if the AI doesn't want to give away its plan, it could queue a message for when it's safe or when it's eliminated to make it clear that there was a plan and not just random stupidity.



One example of this is a behavior I hadn't noticed previous to the Guardians patch, whereby the AI will penetrate deep into the territory of a warring faction without taking the cities it comes across. I'm not sure what it's doing-- is it trying to reach the capital intact? This behavior looks really silly when I see it in action, because the AI is marching right past ungarrisoned cities that it could take without even taking damage-- sometimes, they're cities where there are roads, and taking them would actually speed the army toward the capital rather than slow it. In any case, this is another good situation for some flavor text to give some information about the AI's plan, because all too often it just looks mindless.

natev wrote:
One example of this is a behavior I hadn't noticed previous to the Guardians patch, whereby the AI will penetrate deep into the territory of a warring faction without taking the cities it comes across. I'm not sure what it's doing-- is it trying to reach the capital intact? This behavior looks really silly when I see it in action, because the AI is marching right past ungarrisoned cities that it could take without even taking damage-- sometimes, they're cities where there are roads, and taking them would actually speed the army toward the capital rather than slow it. In any case, this is another good situation for some flavor text to give some information about the AI's plan, because all too often it just looks mindless.

I'm not against the idea of more flavor text to convey the AI's intentions (where appropriate). I still worry that when the AI looks like it is doing something dumb, even with flavor text, if it works to the disadvantage of the AI, it is still technically something dumb. Ideally, Id' like to have less moments where this is occurring period.



The majority of the time, doing what a competent player would do would be the smart thing. If the AI is not doing that, even pretty words won't make it more challenging.



In your example, there is very little that the AI could have been doing that makes sense unless it was somehow worried about expansion disapproval. In which case, how do you win a war without taking cities? Does the AI know how to try to intimidate another player by moving a powerful army inside its territory?

Just playing Devil´s advocate here, because I know in the majority of times the more cities the merrier - but if I was a cold calculating machine, I would probably always go for the least amount of conquered cities which lead me to an elimination victory. If concerns like trade routes are low on my priorities (because im already on terrain which gives plenty science and dust), I would probably leave small cities to be a nuissance for my opponent, not for me. In fact, I do that.



Flavour text regarding that would funny though - "You can keep the villages"

Slashman, I can't say exactly why the AI is coded to behave like this, but my guess is that there is some kind of reasoning involved that is sometimes appropriate, but not appropriate to the situations that the AI finds itself in when playing me. For instance, if it assumes that cities are protected and that it will take damage on cities, then it would be perfectly appropriate for AI to ignore intervening cities to reach the capital. Giving some feedback as to the AI's reasoning isn't just the difference between braindead and misguided behavior-- it's also important for getting feedback from players as to how that behavior could be improved. (In this case: factor cities' garrisons, roads into the decisions to ignore them on a march to the capital.)



Now, I agree, making smart AI is more important than making communicative AI. But wilbefast seems like he or she wants to pick the lowest hanging fruit first: if the AI is doing absolutely anything smart that's unrecognized, even potentially smart, get the player to recognize that the behavior is more than random chance.

natev wrote:
Now, I agree, making smart AI is more important than making communicative AI. But wilbefast seems like he or she wants to pick the lowest hanging fruit first: if the AI is doing absolutely anything smart that's unrecognized, even potentially smart, get the player to recognize that the behavior is more than random chance.

No argument from me on the low hanging fruit angle. I just wanted to make sure that we don't pat the AI on the back for doing something that is detrimental even if it is perfectly in line with its programming. Well specifically, don't pat the coders on their backs.

wilbefast wrote:
Manu made note of this in his original list but I wasn't able to find where he was quoting from - thanks for bringing it up again so I can more officially add this to the list. Philosophical question though: should the AI know that it's in a game with specific victory conditions? By this I mean: is it "realistic" if an AI sacrifices its military or economic integrity because it knows that the end-game is coming up?

I think people's complaints in this thread are largely that AI is not competent, and the corrections (cheats) to improve its competency lead to play very unlike playing against another human. These people are looking for fairer play, not for play with a stronger narrative. (People outside the Amplitude forums, and Amplitude itself, might be looking for something different, of course.)



The fewer things that the AI knows, checks, guesses, the less competent it's going to be as compared to a human. That includes knowledge of the approaching end of the game. It is much easier to handicap a competent AI than it is to start with a handicapped AI and try to make it stronger.



The way I see it, AI needs to take into account every piece of information a player does to even approach competency, because what it does with that information is necessarily going to be weak and formulaic (and likely random). The AI absolutely should know about approaching victory conditions as much as any human player could. After it has this knowledge, easier difficulties can be gimped to "forget" it. The highest difficulty AI should not make sacrifices to its game in the name of narrative. The easier difficulties could. On normal difficulty, I could easily see an AI saying, "Hey, a win for a faction that I like is a win for me."

natev wrote:
I think people's complaints in this thread are largely that AI is not competent, and the corrections (cheats) to improve its competency lead to play very unlike playing against another human. These people are looking for fairer play, not for play with a stronger narrative. (People outside the Amplitude forums, and Amplitude itself, might be looking for something different, of course.)



The fewer things that the AI knows, checks, guesses, the less competent it's going to be as compared to a human. That includes knowledge of the approaching end of the game. It is much easier to handicap a competent AI than it is to start with a handicapped AI and try to make it stronger.



The way I see it, AI needs to take into account every piece of information a player does to even approach competency, because what it does with that information is necessarily going to be weak and formulaic (and likely random). The AI absolutely should know about approaching victory conditions as much as any human player could. After it has this knowledge, easier difficulties can be gimped to "forget" it. The highest difficulty AI should not make sacrifices to its game in the name of narrative. The easier difficulties could. On normal difficulty, I could easily see an AI saying, "Hey, a win for a faction that I like is a win for me."

+1 to the "top/bottom" design.



It´s all much easier said than done; but within possibility, make the hardest levels as human-like as possible and then dumb it down as difficulty lowers by restricting its access to the most efficient algorithms.

BPrado wrote:
+1 to the "top/bottom" design.



It´s all much easier said than done; but within possibility, make the hardest levels as human-like as possible and then dumb it down as difficulty lowers by restricting its access to the most efficient algorithms.

I agree here and with natev as well. This is what Age of Wonders and Fallen Enchantress: Legendary Heroes does. Make the strongest AI possible without bonuses as the middle ground level and then lower levels add handicaps to AI 'thinking'. Higher levels from there add escalating bonuses. That makes the most sense to me.

natev wrote:
Just thinking a little more about it-- wilbefast, you mentioned that sometimes the AI seems dumb, and you look in its innards and it's actually being smart. It sounds like what you're saying is it'd be smart to add flavor text to make the player understand what the thought process is. Even if the AI doesn't want to give away its plan, it could queue a message for when it's safe or when it's eliminated to make it clear that there was a plan and not just random stupidity.



One example of this is a behavior I hadn't noticed previous to the Guardians patch, whereby the AI will penetrate deep into the territory of a warring faction without taking the cities it comes across. I'm not sure what it's doing-- is it trying to reach the capital intact? This behavior looks really silly when I see it in action, because the AI is marching right past ungarrisoned cities that it could take without even taking damage-- sometimes, they're cities where there are roads, and taking them would actually speed the army toward the capital rather than slow it. In any case, this is another good situation for some flavor text to give some information about the AI's plan, because all too often it just looks mindless.

I'm not going to say that sometimes it isn't just being dumb but yes there are certain circumstances where it seems a lot dumber than it really is!

natev wrote:
I think people's complaints in this thread are largely that AI is not competent, and the corrections (cheats) to improve its competency lead to play very unlike playing against another human. These people are looking for fairer play, not for play with a stronger narrative. (People outside the Amplitude forums, and Amplitude itself, might be looking for something different, of course.)



The fewer things that the AI knows, checks, guesses, the less competent it's going to be as compared to a human. That includes knowledge of the approaching end of the game. It is much easier to handicap a competent AI than it is to start with a handicapped AI and try to make it stronger.



The way I see it, AI needs to take into account every piece of information a player does to even approach competency, because what it does with that information is necessarily going to be weak and formulaic (and likely random). The AI absolutely should know about approaching victory conditions as much as any human player could. After it has this knowledge, easier difficulties can be gimped to "forget" it. The highest difficulty AI should not make sacrifices to its game in the name of narrative. The easier difficulties could. On normal difficulty, I could easily see an AI saying, "Hey, a win for a faction that I like is a win for me."

Best summary of the issues I have read.

Righto - so knowledge of encroaching victory seems to have consensus

natev wrote:
I think people's complaints in this thread are largely that AI is not competent, and the corrections (cheats) to improve its competency lead to play very unlike playing against another human. These people are looking for fairer play, not for play with a stronger narrative. (People outside the Amplitude forums, and Amplitude itself, might be looking for something different, of course.)



The fewer things that the AI knows, checks, guesses, the less competent it's going to be as compared to a human. That includes knowledge of the approaching end of the game. It is much easier to handicap a competent AI than it is to start with a handicapped AI and try to make it stronger.



The way I see it, AI needs to take into account every piece of information a player does to even approach competency, because what it does with that information is necessarily going to be weak and formulaic (and likely random). The AI absolutely should know about approaching victory conditions as much as any human player could. After it has this knowledge, easier difficulties can be gimped to "forget" it. The highest difficulty AI should not make sacrifices to its game in the name of narrative. The easier difficulties could. On normal difficulty, I could easily see an AI saying, "Hey, a win for a faction that I like is a win for me."

So giving the AI a "crystal ball" on higher difficulties? Letting it see things it shouldn't be able to like how many armies you have and where they are, whether you're building the wonder it wants to, when winter will come and so on? Philosophical question number 2: does a mildly psychic AI feel more or less fair to play against than one which is inexplicably slightly richer than you are, but no-less well-informed (actually slightly less well-informed in this case)?

I don't think that means giving the AI more info than the player does, just access to and consideration of the same information.



A relevant example is the decision on whether or not an Empire thinks it can win a fight. Is the AI considering things like relative Initiative (I'm going to lose a few units before I attack), reinforcements (more units than the target army) and special abilities (hero with Heal is going to negate an attack each turn)?

wilbefast wrote:
Righto - so knowledge of encroaching victory seems to have consensus







So giving the AI a "crystal ball" on higher difficulties? Letting it see things it shouldn't be able to like how many armies you have and where they are, whether you're building the wonder it wants to, when winter will come and so on? Philosophical question number 2: does a mildly psychic AI feel more or less fair to play against than one which is inexplicably slightly richer than you are, but no-less well-informed (actually slightly less well-informed in this case)?

But why give it knowledge that the players doesn´t have? Why for example should this AI know "where" your armies are? Can´t it just have access to the Military score like a player does and react to that?



I think it depends on the kind of extra information. If the AI knows the composition of your armies (which a player could by clicking any Army in sight), then it can retrofit its own armies on a good Meta; but if the also AI knows where your armies are at all times, and it actually uses the information, then mostly any bonus to industry would probably make wars terribly frustrating.





edit: Ninja´d by prop, but yeah, that´s another good example. If an AI knows exactly where your armies are, it can´t fall for reinforcement traps. I think it should be passive of falling for this kind of thing on lower difficulties (as a bad player is less able to predict the outcome of a fight through the battle overview), but higher difficulties shouldn´t.



In this case, the omniscience would probably emulate the actual behavior, but isn´t there a way to make the AI react to the battle, and not to the armies´ positioning?

wilbefast wrote:
So giving the AI a "crystal ball" on higher difficulties? Letting it see things it shouldn't be able to like how many armies you have and where they are, whether you're building the wonder it wants to, when winter will come and so on?

Just all of the information that the player has (Propbuddha understood correctly.) But I'll try to address the next question if you want:

Philosophical question number 2: does a mildly psychic AI feel more or less fair to play against than one which is inexplicably slightly richer than you are, but no-less well-informed (actually slightly less well-informed in this case)?

When I know I'm playing against psychic AI, it makes me irritated, because it rules out creative strategies I might come up with where I try to fool the AI. Psychic AI might work in some games where I never realize it's happening, but it probably takes some work to achieve that level of subtlety.



Before investing time to give the AI more information than the player has, give it as much information as the player has. Giving psychic abilities to the current AI is not going to make it play more like a human would.

natev wrote:
Before investing time to give the AI more information than the player has, give it as much information as the player has. Giving psychic abilities to the current AI is not going to make it play more like a human would.

Exactly this. Approaching end of game, knowledge of the composition of visible player armies, knowledge of its factions' unique strengths and weaknesses. These are things that we want the AI to know.



Knowing where all armies are (both visible and invisible), knowing what other players are building, knowing where all player territories are and how far they extend without scouting. These are things that the AI should not know or be able to do.