Query Regarding AI

[sravanind]

Hi,
[this query is for David, but anybody is welcome to answer]

I'm studying/analyzing AI in few strategy games like AOE, AOM.
I'm trying to analyze AI of Capitalism 2.
it'll be great help if anyone can share some points about AI in Capitalism
like
1) what approach is used for designing AI? I mean does it use fuzzy logic or Finite State Machine or any other approach
2) what is AI competitors goal
3) how AI competitors achieve their goal
4) how aggressiveness effects AI strategy

Thanks in advance,
Sravan

[counting]

Actually I am also very curious about these questions.

IMO, since I can predict quite a few AI behaviors in the game, I would say it's fairly deterministic, hence I suspect it might be certain weight average and threshold triggered conditions. And AI corporations have "personalities", where different focused AIs behave drastically different but always the same way within each focus (although they can change personality half way), and I've observed weird AI behaviors in off boundary situation, and it always get stuck (or simply do nothing) facing the same dilemma (especially very late decades to a century into the game, where all markets are fully saturated).

I am curious about whether FuSM is part of the design though, I notice AIs in CapLab have tolerance for certain price range when choosing suppliers. And it's related to how aggressive AI is set. But the most noticeable difference in AI aggressiveness is its "timing". At lower aggressiveness, AI take longer to take actions (like tuning price, or venturing into new businesses).

As for goals, I have a very strong feeling that AI mostly operate on the firm level, instead of corporate level. A lot of time AI behave very aggressively and well in individual firm, but not so well when aggregated together. It doesn't have a higher level of planning like most human players can, like it has tunnel vision (what I like to call suicidal tendency), but most gaming AIs suffer similar issues. And for the most part, if you give CapLab AI enough starting funds and proper start conditions, it can behave quite well to make the game challenge enough.

[sravanind]

Thanks Counting, for sharing your observations.

IMO, since I can predict quite a few AI behaviors in the game, I would say it's fairly deterministic, hence I suspect it might be certain weight average and threshold triggered conditions. And AI corporations have "personalities", where different focused AIs behave drastically different but always the same way within each focus (although they can change personality half way), and I've observed weird AI behaviors in off boundary situation, and it always get stuck (or simply do nothing) facing the same dilemma (especially very late decades to a century into the game, where all markets are fully saturated).

yes, AI is stuck after a while into the game. may be boundaries are hard limits. after reaching those limits AI just stuck.

I am curious about whether FuSM is part of the design though, I notice AIs in CapLab have tolerance for certain price range when choosing suppliers. And it's related to how aggressive AI is set. But the most noticeable difference in AI aggressiveness is its "timing". At lower aggressiveness, AI take longer to take actions (like tuning price, or venturing into new businesses).

I'm also curious about FuSM. if FuSM is involved in AI design, then AI wouldn't be stuck after some time.
as per me, AI aggressiveness should be in-terms of starting new factories or farming or may be retails stores, which gives tough competition to user.

As for goals, I have a very strong feeling that AI mostly operate on the firm level, instead of corporate level. A lot of time AI behave very aggressively and well in individual firm, but not so well when aggregated together. It doesn't have a higher level of planning like most human players can, like it has tunnel vision (what I like to call suicidal tendency), but most gaming AIs suffer similar issues. And for the most part, if you give CapLab AI enough starting funds and proper start conditions, it can behave quite well to make the game challenge enough.

so, AI doesn't have long-term goal set? it behaves aggressive in parts?
this is most interesting part to know in the game, what exactly is the AI's target?

[counting]

yes, AI is stuck after a while into the game. may be boundaries are hard limits. after reaching those limits AI just stuck.

Some situations where AI gets stuck are pretty weird though. Like AI will always try the large factory when they have enough found, even when there's only small demand (even though large factory have higher production-to-wage ratio, but its high base overhead will cost more than less efficient middle/small factory, and most of the time small/medium locate factory is enough). Human players can detect this kind of issues quite easy with financial reports. (Even though AI will try to put these factories on-sale, if they stay in negative profit for too long). On the other hand, AI will usually bottleneck its own production, especially with semi-products. Even for the most aggressive AI setting, it still doesn't scale up well, where clearly there are still blank market to expand downstream production.

as per me, AI aggressiveness should be in-terms of starting new factories or farming or may be retails stores, which gives tough competition to user.

Personally, I find even the most aggressive AI setting too passive. And a lot of time, it just feels like specifically targeting human player instead of maximize its own profit. Especially when human players venture into new markets, and suddenly AIs begin to join the competition right (a lot of time, they don't even have the tech or brand to fight). And if reloading a previous save before entering, the blank market will stay blank. (Although I can not rule out, it's due to human entering increases the city over-all rating, create a larger market, thus give AI incentive to follow)

so, AI doesn't have long-term goal set? it behaves aggressive in parts?
this is most interesting part to know in the game, what exactly is the AI's target?

I am not sure how to define long-term goal. How do humans determine long-term goal?

I think humans divide tasks subconsciously, like this is early game, the company needs to survive, the task is to meet break-even point, in order to do that, the first sub-task is to analysis all the AIs' expertise, then cities population/wage rate, then different product ratings, seaport goods, etc. I don't think AI does such planning ahead, but just follow per-defined scripts, do this and do that base on assigned personality (especially obvious if they are real estate AIs, or stock focused AI). When entering mid-game, human usually try to analyze what kind of products share semi/raw and try to expand toward those products, but manufacturing AIs seems to venture into one product at a time, regardless product connectivity, but just choose those products that either are their expertise first, and then low over-all city rating products, so on. In the late game, AI normally just feel like an aggregate firms, when you attack one of its market, it only defense "passively" with price war without strategies. It doesn't know how to fix massive negative corporate brand loyalty, or try to find other profit margin when the original profit base is lost. And their "strategy" of purchasing tech from each other is very confusing, like food expertise AIs buy cosmetic tech, household product AI buy automobile tech, and most of the times they don't use any of these purchased techs, however from time to time, a lucky tech-focused AI will gather the right combination, and turns into diversify AIs, then takes over majority products. And this is the interesting part, with enough AIs, they behave some what intelligently like swarm intelligence, individually it might be vulnerable to human strategies, but collectively they become harder and harder to beat (like some might go bankrupt due to human tactics, but another one with more capital will be back to take its place. With swapping techs, they catch up very quickly).

[sravanind]

wow, that is great analysis of the game.

Like AI will always try the large factory when they have enough found, even when there's only small demand (even though large factory have higher production-to-wage ratio, but its high base overhead will cost more than less efficient middle/small factory, and most of the time small/medium locate factory is enough)

now I see, why so many "For Sale" of AI players. I think AI may not be considering production-to-wage ratio and demand or maybe it's going for trial and error.

I am not sure how to define long-term goal. How do humans determine long-term goal?

when I start a game, may be I want to dominate market in particular domain or may be want to be the richest

I think humans divide tasks subconsciously, like this is early game, the company needs to survive, the task is to meet break-even point, in order to do that, the first sub-task is to analysis all the AIs' expertise, then cities population/wage rate, then different product ratings, seaport goods, etc. I don't think AI does such planning ahead, but just follow per-defined scripts, do this and do that base on assigned personality (especially obvious if they are real estate AIs, or stock focused AI).

yes, you are absolutely correct. we divide tasks and first priority will be to establish a safe zone and then venture further towards the goal.
This long-term goal seems non-existent for AI.

their "strategy" of purchasing tech from each other is very confusing, like food expertise AIs buy cosmetic tech, household product AI buy automobile tech, and most of the times they don't use any of these purchased techs, however from time to time, a lucky tech-focused AI will gather the right combination, and turns into diversify AIs, then takes over majority products. And this is the interesting part, with enough AIs, they behave some what intelligently like swarm intelligence, individually it might be vulnerable to human strategies, but collectively they become harder and harder to beat (like some might go bankrupt due to human tactics, but another one with more capital will be back to take its place. With swapping techs, they catch up very quickly).

I think AI is trying to distribute tech among themselves. it may be rare that one or more AIs gather right combination. even if they act as swarm it's late in the game to give competition to human player. at-least AI should collaborate and conspire against the player and push one AI player to give competition to human player from the beginning.

if reloading a previous save before entering, the blank market will stay blank.

this may be file save/load problem. may be AI not getting all the input it required to restore it's position to where it was last time when saved.

[counting]

I am not sure how to define long-term goal. How do humans determine long-term goal?

when I start a game, may be I want to dominate market in particular domain or may be want to be the richest

For the most part, I believe the game AI doesn't actually have a goal orientated design, but more like go with the flow (adaptive design). And the adaptive mechanism mostly operates at firm level. , since these AIs need to be able to function in both sandbox/custom games, and scenarios. The best way I can describe these AIs is that they are like zombies, you can hack and slash them in any parts, but they keep going until you can fully dismembered them (take away all their profit margin), or go for their heads (hostile take over)

I think AI is trying to distribute tech among themselves. it may be rare that one or more AIs gather right combination. even if they act as swarm it's late in the game to give competition to human player. at-least AI should collaborate and conspire against the player and push one AI player to give competition to human player from the beginning.

I would like to see that too. Swam intelligence scripts shouldn't be too hard to implement, and they could theoretically create many more unpredictable strategies combination, then pre-defined scripts. But this requires certain game theory mechanism in place for AIs to "understand" the co-op and how to distribute "goal points" among them.

this may be file save/load problem. may be AI not getting all the input it required to restore it's position to where it was last time when saved.

There are lots of possibilities though, even perhaps AI uses some sort of random seed.

[sravanind]

For the most part, I believe the game AI doesn't actually have a goal orientated design, but more like go with the flow (adaptive design). And the adaptive mechanism mostly operates at firm level. , since these AIs need to be able to function in both sandbox/custom games, and scenarios. The best way I can describe these AIs is that they are like zombies, you can hack and slash them in any parts, but they keep going until you can fully dismembered them (take away all their profit margin), or go for their heads (hostile take over)

in games like AOE, AI tries to attack the player based on difficulty level. which forces player to ready defenses early on and increase resource gathering and build military units etc.,
seems like here player has to take-on AI and get it out of the game.

I would like to see that too. Swam intelligence scripts shouldn't be too hard to implement, and they could theoretically create many more unpredictable strategies combination, then pre-defined scripts. But this requires certain game theory mechanism in place for AIs to "understand" the co-op and how to distribute "goal points" among them.

I think script may not be sufficient. internal implementation may need a change for this kind of AI approach.