wilbefast wrote:
Great post - very in-depth I've book-marked it for later reference and will probably pull a few things out of here!



For the record EL's GO list was essentially:

[LIST=1]

Optimise my economy/happiness/resource production within my current borders

Expand if happiness allows for it and there are suitable neighbouring regions to colonise

Make sure I am not in danger of being attacked and wiped out and do my best to win current war(s)

Attack (weaker) opponents, especially opponents with a higher science or economic score

[/LIST]

If this i true - or at least very close to be - I believe the priorities need to be changed. What makes the strategic AI bad in EL is in my opinion primarily how military weak it is. For a human player it is very easy to exploit an AI that first seeks to build up economy and found new cities, before protecting what it already has. What does another small city or an economic structure do for me if I am attacked and basically wiped out in a few turns? I believe there could be a pretty significant performances gain in just redefining the priorities of the AIs GO List. Obviously the priorities may be a bit different during the first 20 turns or so, but the AI should quite early in the game make sure it is not an easy target for any nearby enemy. Unless something is changed here, it will forever remain very easy for a human to exploit AI controlled factions by just focusing on military might.

wilbefast wrote:


So what's missing really is the focus on victories: both the AI's own victory and those of other players. We figured the AI should not know it is in a game, meaning that it doesn't behave as though the game is going to end on turn 300 or when another empire reaches X dust. This felt like it would be cheesy to us. Perhaps we can find some sort of middle-ground though for ES2. Thoughts?

I disagree. I play for the challenge. Having an AI that does not care about winning or losing, that basically plays blindfolded and then once in a while just happens to stumble into a victory is not challenging, nor does it make any sense in the context of the game. The AI is already handicapped compared to a human. There is no need to deliberately make it not know the rules/goals of the game.



Such an AI is very easy to exploit, and quickly becomes repetitive and boring. I understand that you want "role playing" AIs to add immersion, but I don't think role replaying and competitiveness is mutually exclusive. Basically you make sure the AI plays to win, which also means it should focus on weakening the strongest competitors. It should, if possibly, use military might to exploit weaker opponents and/or stop competitors from winning the game. It should also seek to form alliances with empires that do not pose a threat (much lower global score/not close to a victory) in order to take on the stronger opponents. Let's say you have a game with 4 empires:



Red - global score 2000, close to economic victory

Blue - global score 1600

Green - global score 600

Orange - global score 600.



- Red should seek to defend it's good position and secure the Economic victory, while potentially conquer Green and Orange (which wouldn't be too distracting from the overall goal of Economic Vic)

- Blue borders Red and Green. It should seek to ally Orange to take on Red, and as a side project at some point try to conquer Green.

- Green and Orange should both switch focus to a peaceful victory condition based on faction/regions/techs, and at the same time seek alliances with basically anyone except Red. Can they form an alliance strong enough, they should take on Red.



The key to keeping this kind of stuff "in role" is to add good flavor texts and explanations during diplomatic negotiations. If explained well, I even think it would add a lot. As an example, Blue and Orange when declaring war on Red could explain that the endless greed of Red is killing the economies of all other nations, and that the population outside of Reds empire is living in poverty. That they will not stand idle by when Red monopolizes the dust production, and alone reaps the rewards. It would make sense both from a gameplay and lorewise point of view. (I am sure most people could come up with something better, but you get the idea).



Having a competent AI should be the main priority. A competent AI can still be a pretty good role player, as long as all the factions are designed in a way that allows for proactive play that does not conflict with lore.



EDIT: Also, playing to win peacefully does in no way conflict with lore. Gathering as much dust as possible makes perfect sense Roving Clans and Broken Lords in particular. The same can be said about science and Vaulters/Arden Mages and diplomacy points for Drakken. Also remember that basically all factions can benefit from science, dust and diplomacy, and even if they are not particularly designed with an economic victory in mind, it does make sense to accumulate all the dust they can if they are given regions that allow it. Both in the context of the game or lore.

darkath wrote:
This happens a lot in Civilization, you don't understand why Ghandi is suddenly sending nukes, so ultimately you don't bother trying to do diplomacy with the AI anymore since you can't figure out what works and what doesn't (there is so many possible situation that ultimately it's impossible by pure trial and error)

I agree with you on the part about transparency/immersion, don´t get me wrong.



But, sorry, I´m going to be extremely annoying here and point out that - no, at least Civ5 does not act randomly. The Gandhi/nukes thing is a joke that´s completely managable. Diplomacy in Civ5 is extremely detailed, heavily based on realist politics and very "predictable". I´m more than happy to explain privately mostly any scenario you might not understand.



It might look confusing at first because everything counts - It´s different you have one unit on their borders or if you have 4, it´s different if you´re their friend and you´re weak and if you´re their friend and you´re strong, it´s different if you´re winning and you´re first and if you´re winning and you´re third, it´s different if you try to be friend with everyone and if you actively choose your friends and enemies, it´s different if you´re friends with someone and both of you have a religion or if only one of you do. Diplomacy in civ5 cannot be currently compared to Diplomacy in EL.





edit:



I agree with clykke up here in terms of challenge.

clykke wrote:


Having a competent AI should be the main priority. A competent AI can still be a pretty good role player, as long as all the factions are designed in a way that allows for proactive play that does not conflict with lore.

I just feel that having a competent AI is the main priority of an offline game, not necessarily MP, and I hope Amplitude themselves correct if I´m wrong, but my feeling is that Endless Legend (at least) is focused on MP, on which case the Mechanics are the main priority.

BPrado wrote:
I agree with you on the part about transparency/immersion, don´t get me wrong.



But I´m going to be extremely annoying and point out that - no, at least Civ5 does not act randomly. The Gandhi/nukes thing is a joke that´s completely managable. Diplomacy in Civ5 is extremely detailed, heavily based on realist politics and very "predictable". I´m more than happy to explain privately mostly any scenario you might not understand.



It might look confusing at first because everything counts - It´s different you have one unit on their borders or if you have 4, it´s different if you´re their friend and you´re weak and if you´re their friend and you´re strong, it´s different if you´re winning and you´re first and if you´re winning and you´re third, it´s different if you try to be friend with everyone and if you actively choose your friends and enemies, it´s different if you´re friends with someone and both of you have a religion or if only one of you do. Diplomacy in civ5 cannot be currently compared to Diplomacy in EL.

It's been a long time since i've played CIV but what i meant is as far as i remember you won't have a detail of why Ghandi (or anyone else) suddenly hates you (or may be i'm thinking of CIV4 here).



But i agree with you. The AI is not actually acting randomly, what i meant is that if you conceal the information about "the why", it could confuse the player that would then think the AI acts randomly.

BPrado wrote:


I just feel that having a competent AI is the main priority of an offline game, not necessarily MP, and I hope Amplitude themselves correct if I´m wrong, but my feeling is that Endless Legend (at least) is focused on MP, on which case the Mechanics are the main priority.

With main priority I mean between having a competent AI and a role playing AI. Not for the game as a whole

wilbefast wrote:
So what's missing really is the focus on victories: both the AI's own victory and those of other players. We figured the AI should not know it is in a game, meaning that it doesn't behave as though the game is going to end on turn 300 or when another empire reaches X dust. This felt like it would be cheesy to us. Perhaps we can find some sort of middle-ground though for ES2. Thoughts?

There is no cheesiness when the AI is armed with the same knowledge that the player is armed with. It is, in fact, what should be the norm. There is no advantage to having a less aware AI unless the things that the AI is aware of are things that the player is not. Like, knowing how much dust each faction has or something omniscient. Otherwise, the AI should have every advantage that the player has in terms of game knowledge.



Also +1 to clykke for his post on challenge.

In a similar vein, what about the system AI for players to use?



What about adding an optional 'allow terraforming' thing so it doesn't go rampant with planet type changing, especially if the faction is like Sowers or Amoeba who have a researchable boost to a specific planet type?

What about publishing AI API and publishing code of baseline AI?

In many games community developers managed to create(improve baseline or even just tweaks parameters) AI that was much better than baseline.

ES AI create so terrible designs that it's not possible to measure it general AI is capable of doing anything. (just noticed mod that make attempt to solve this issue, maybe it wouldn't allow player to rollover endless AI).



Examples of community enhanced AI:

TW series - Darth Mod(both battle and campaign AI) and Radious AI mod(campaign AI). Thay acomplished quite a lot considering that there's no AI API and there's no support from CA for their efforts.

Cities Skylines - Trafic++ - this mod recreate agent AI of vechicles from scratch. AI API is avaible, however it's not documented. this community AI mod fixed nearly all issues with baselane AI.

Cosmic Supremacy - player could write limited AI for their fleets(called admirals) and systems(called governors). Some of this scripts are very sofisticated and could handle perfectly nearly whole micromanagemenet and even handle empiere when player is away for several turns.

The tweaks from Darth Mod are quite impressive, but I guess that´s contrasted with the godawful combat AI the series has.



I feel a little bad throwing these bad adjectives around on such ambitious projects, but not too bad because mods like these prove that developers can come up with better routines if they´re not smashed by publishers.

The most important thing for believable AI is to manifest as a character. EL is OK in showing what concerns AI and why he hates or likes you. But it does very poor job in showing what his priorities are. Does this guy remembers you've wronged him more intensely than the other? Does this guy value agreements or not? They need to talk more. EL guys just say the agreement is no longer interesting to them when they break it but they can do a better job about it - "We don't want this trade agreement cause it enriches you more than us", "We don't like your warmongering and don't want to fuel your warmachine" etc.



Playing against enemies who have specific geopolitical agenda is what I want. I don't like seeing them build their cities sub-optimally or missing attack opportunities as long as they feel rational. Let them be stupid, but make them feel sentient.

Mer wrote:
What about publishing AI API and publishing code of baseline AI?

Interesting idea, and there were several academics asking for open API for game AI at Nucl.ai. I'll discuss it with the team



What would constitute a "baseline AI" for you? We'd rather not give away all our technology (ie. open-source our AI code) and given the choice it's probably better to spend the time making said technology better rather than writing a new, simplistic version as an example for modders.

There are several possibilities depending how much would you like to share, more you share then faster and better results you would get.

Option A - full share - you give away technology, yet you would get best community results.



Option B - access to structure of AI and API for compiled classes(ex. cities skylines). You keep your code, however your AI could be used as base for development. It require quite a lot of documentation and clear structure. (only sub-packages API is required to be published, sub-packages private classes could stay secret)



Option C - only game API for AI and basic AI, as simple as possible yet capable to alone play game(if given enough time it should 1. colonize hole map, 2. research all technologies, 3(optional). conquer passive opponent. It's always hardest to start.

wilbefast wrote:
Interesting idea, and there were several academics asking for open API for game AI at Nucl.ai. I'll discuss it with the team



What would constitute a "baseline AI" for you? We'd rather not give away all our technology (ie. open-source of AI code) and given the choice it's probably better to spend the time making said technology better rather than writing a new, simplistic version as an example for modders.

Just being able to tweak different values and priorities would be fantastic start.



However a full blown AI API would be absolutely brilliant.



EDIT: I'd kill to get the chance to actually see and change the AI code, and then test it. I know it won't happen, but just mentioning it as a possibility gets me all excited. I'd probably forget to eat the first few weeks...

Since you have a bunch of experienced people working on it, I feel I may be stating the obvious, but I'll share my 2 cents as well.



This is full of conjectures/assumption about the game's design, mind you.



I think you should be ambitious in how many AIs you should allow players to add to the game. I'd say anything from 64 to 1024 is a good value (256 is good enough).



The AI would be a single step in the game process. Being a turn-based game, each AI would process the game state - or preferably as much as it knows about the game state, if it is not omniscient - serially. It would be crucial then to make the AI thought proces withing a single turn to be multithreaded for performance purposes. The AI would manage fleets, the map, resources, look out for enemies, choose where to invest FIDSI, where to go, manage diplomacy. Those are all different problems and need not be serialized inside one single AI 'subturn'.



I imagine that each factions decision process - their subturn inside one game turn - would have a few stages.



Explorer thread - looks at the map and where to go, proposes using X resources (fleets probably) to expand its 'domain problem'. May suggest

creating more exploration fleets; if there are known hostiles or neutrals, may suggest militarized fleets as well.

Fleet exploration/border protection

Colonization/lane blocking to secure future planet expansion



Economic thread - looks at all owned/known planets, proposes:

improvements to owned planets

science research to foster economy

planets to colonize

planets to conquer

peace/trade with other factions

scrap fleets/planets (if budget really imbalanced)



Military thread - looks at current fleet and compares to what is known about other factions, how much it feels it needs to defend the system, tech levels and proposes:

Military tech

Fleet production

Planetary military improvements

Fleet repositioning

War/truce/peace



Siege thread - tactical close-up look to invade a specific planet/system. Asks for fleets/troops and such.



Diplomatic thread - compares different broad (known) aspects of enemies and intelligence information to suggest best stances towards other factions.

War/truce/neutral/friends/trade/alliance



Grand goals: each faction would get all the possible victory outcomes arranged randomly in a priority order on each game. They would probably pursue the one at the top, but depending on the environmental/economic/diplomatic factors, may pursue others, usually tending to make decisions which maycollaterally help achieve the greatest goal. The grand goals are a set of variables created at the beginning of each game and don't change.



All those proposals would be weighted numbers, and a second level of the AI decision-making which does the grand strategy would look at the proposals and tweak the strategy turn-by-turn, also taking into account:



Its own personality

The grand goals



There should be some overlap between each thread AI, with different threads possibly suggesting similar things to do, which would add them weight.



Each thread should have a 'governor profile', which would give them a small bit of personality, leaning towards some suggestion more than others. This would be changed per-game.

THis would add unpredictability to the enemy faction AIs, since even though each faction has a more or less defined personality, it would still have a gradation inside that personality.



The game will have a single 'true state', but as each faction explores, it will have a bunch of photographies of the game state as it was back then. They should make decisions based on what they know. This should be pursued as much as possible, anc cheat mechanics used only as a last resort. I'd prefer to have cheat mechanics being indirect measures instead of direct ones.



AI cheating - or not



If I were to do the AI cheating, I'd have it as something integral to game lore. Suppose there is a super-sentient transcendental alien race (Endless? the Endless' creators? - Endless Devs teehee) which know everything, but is not interested in cosmic exploration/expansion/domination.

It would be a separate thread as well, to be run before each turn. I would have it as a super-spy which gives random bits of information to the AIs 'out of goodwill', updating and expanding the AI's photography of the game state to a more complete picture, so that it can make better decisions.

Perhaps even gifting random bits of resources. THis thread could be in charge of the random cosmic events which affect the game, and weight them more in favor of the AIs to help them (if necessary).

It could even be made a neutral power, possibly giving information to the player as well. The difference would be that the parameters - how much, how oftenit gives information - would be tweakable per-faction and could even be reduced to zero, effectively disabling it.

This bias would be tweakable depending on game difficulty, and could even be scaled in the player's favor.

Brazilian_Joe wrote:
I think you should be ambitious in how many AIs you should allow players to add to the game. I'd say anything from 64 to 1024 is a good value (256 is good enough).

256 empires in the same game? O:



I'm afraid this would cause a host of UI and performance issues. The issue isn't address-space, we can spare 10 bytes for identifiers. The issue is that the time taken to calculate the next turn isn't linearly dependant on the number of empires, it's polynomial IIRC (or perhaps exponential). So we're unlikely to have more than in previous Endless games I'm afraid.

Brazilian_Joe wrote:
The AI would be a single step in the game process. Being a turn-based game, each AI would process the game state - or preferably as much as it knows about the game state, if it is not omniscient - serially.

I'd like us to focus on feature requests here rather than implementation details - figuring out what the best architecture is is our job

Brazilian_Joe wrote:
If I were to do the AI cheating, I'd have it as something integral to game lore. Suppose there is a super-sentient transcendental alien race (Endless? the Endless' creators? - Endless Devs teehee) which know everything, but is not interested in cosmic exploration/expansion/domination.

Interesting idea! We'll have to see what Jeff thinks of course.

clykke wrote:
Just being able to tweak different values and priorities would be fantastic start.

There are a certain amount of things that can be tweaked in Steam\steamapps\common\Endless Space\Public\AI and \Steam\steamapps\common\Endless Legend\Public\AI respectively - best refer to the corresponding mod forum to find out what affects what. We're likely to have parameters made public in ES2 as well.



[HR][/HR]



With regards to the API: yes, it's an interesting idea - I'd like to stress though that my saying "that's an interesting idea" is by no means a promise that we will have any such thing

wilbefast wrote:
I'd like us to focus on feature requests here rather than implementation details - figuring out what the best architecture is is our job

Random guy from forum knowing nothing about your code and giving you broad advice on architecture is not enough for you?! Gosh, Amplitude just can't be satisfied.



Anyway I think Amplitude is on the right way after EL. You've already made clear what AI thinks about you. You're going to add character to, er, characters I've heard. My minor issue is with consistency and evaluations - I still don't know why AI values Dust so much more than other resources and why he spends influence on destroying agreements even though he doesn't seem to be gaining anything from it or desire to attack me. But it's much better than even release EL. And they still screw up city building I think.

From my point of view (single and casual player) I really don’t care if the AI cheats or not, I just want an interesting and fun game. If the AI needs to be omnipresent to achieve that, so be it. I really don’t even care if the AI is playing the same game or by the same rules, as long as it’s fun.



I think, and I have no experience in coding what so ever, it would be good for an AI to have a set goal it needs to achieve. Those goals don’t need to be victory conditions.



It could work like this:

At start of the game each AI chooses secretly to the player one random goal (weighted be AI aggression, so warmongers don’t get ‘Get peace’).



Some example goals:

Get rich! – The AI wants to make money. At first it will just research and build an economy, later it will try to conquer the planets with the biggest economy rating. It will never invade planets with negative economy. Ships will be cost efficient.



Get smart! – The AI is going for science victory. It will actively trade for technology and try to conquer planets with science boni. It will try for every research alliance it can get and might pay for it. Ships are always state of the art.



Get lonely! – The AI hates everyone without prejudice. It will trade for weapons, but for nothing else. In war it will always side with the winner first and then turn on them. They will conquer planets, but if it is unlikely to hold them, the will destroy them, if the tech is available.



Get together! – The AI wants to be allied with everyone, but the worst warmongers. To achieve that it will trade, bribe and bully. Whatever it takes. It will always side with the biggest alliance (most factions), not with the strongest.



Get peace(y)! – The AI shuns war. It will never declare war, but their planets will be the best defended in the galaxy. It will be a diplomatic powerhouse and will try to broker peace even between other factions.



Get ‘em all! - The AI loves artifacts (or special resources). It will buy, conquer, loan and pillage them whenever possible. It will go for peaceful variants first, but will use force if necessary. They will starve the galaxy, if not kept in check.



Get ‘em through! – The AI has the knowledge to summon ancient beings from beyond space. All it needs are some rare resources and a couple (ex 7) (randomized) planets to build the constellation to do this. If they succeed the planets AND the faction is destroyed and 7 very powerful space monsters invade the galaxy.



Get in line! – The AI wants to get back to its ancestral home ... or what it believes it is. After the starting position is decided another planet at the other side of the galaxy is chosen. The AI needs an unbroken chain of planets from start to finish. If it succeeds the target planet is transformed into a super planet (depending on the number of planets in the chain) and all other planets get wiped of all colonies, getting free to colonize again. The AI now choses a victory condition and tries to achieve that.



With goals an AI would be able to focus its resources much better I think and thus would be more of a challenge and would give each game some more personality.

Good to know that AI is being seriously worked on. I just hope that your team won't just end up giving cheats to AI and call it a day (many gave up on AI and took an easy road afterall). When I've been playing ES, I usually just end up wrecking AI and starting to cheer for it (against myself).



Cheers.

Hey All:



Just to introduce I was the author of the AI Improvement Mod for Distant Worlds and participated heavily on the forums when AI Improvement was a focus of many patches. Happy to see an AI focused thread here and acknowledgement of the importance of this subject to set the game apart from the competition. I must admit I'm in that group that has been extremely disappointed in the AI of recent 4X Space releases such as Gal Civ 3 (which seems a light-year away from Gal Civ 2) and Stardrive 2.



http://www.matrixgames.com/forums/tm.asp?m=3647528



The area which stands out of the most is that the Endless Legend and Endless Space AI has never been able to build decent ships/units. This is obviously not a new subject but have you identified the root cause of this? From a players perspective the problem has seemed persistent, it didn't matter what patch or what was tweaked. This alone made me question whether Amplitude would really be able to find it's way out of this AI black hole.



Why not apply an expert systems approach where the AI has access to expert designs (i.e. from your best human players) only? The AI needs to build across a range of circumstances and chose what designs to apply but it should choose within that limited range of designs to ensure no weak designs are built. I would suggest seeking input from your VIP's during the Beta and integrating the best of their designs into the game. Ideally have a tracking mechanism to see what your expert players are doing in these key areas and apply that to improve the AI via patches after release also.



The same applies to research orders i.e. consolidate a "base" research order for each race based on expert players and then code variations for specific circumstances / goals. Within those "base" scripts you can of course still capture the personality of each race indeed it's essential to maintain immersion.



I look forward to reading ideas from both the community and developers.

I can vouch for Icemania. He and I did a ES challenge several months ago. How to beat the game on Endless difficulty with the most negative traits. There was a third person too, he's name is WabbaJack. These two guys found ways to beat the AI, without mods or cheating, by just using cheese tactics and minimal/optimal builds. They showed me things that I didn't even know were exploitable after 350+ hours of play. His AI mod for DW:U is a game changer. Kind of like the Long War mod for XCOM:EU/W. He's suggesting some serious ideas that could improve GalCiv3 AI and overall gameplay as well.



He knows what he's about.

Icemania wrote:
Why not apply an expert systems approach where the AI has access to expert designs (i.e. from your best human players) only? The AI needs to build across a range of circumstances and chose what designs to apply but it should choose within that limited range of designs to ensure no weak designs are built. I would suggest seeking input from your VIP's during the Beta and integrating the best of their designs into the game. Ideally have a tracking mechanism to see what your expert players are doing in these key areas and apply that to improve the AI via patches after release also.



The same applies to research orders i.e. consolidate a "base" research order for each race based on expert players and then code variations for specific circumstances / goals. Within those "base" scripts you can of course still capture the personality of each race indeed it's essential to maintain immersion.



I look forward to reading ideas from both the community and developers.

I think it might work much better for tech path than army designs, even because the first heavily affects the latter.



Having an efficient army is relative to your opponents, I don´t think there are specific designs that are absolutely better in Endless Legends (with which many people here will disagree). I think it´s more about having the equipment/dust available when you need them available - there might be timings that are absolutely better. The only constant is that everyone needs strategic equipment.