Iskabis Iskabis

"Incredible" level AI and game balance issues

▼

"Incredible" level AI and game balance issues

from Stardock Forums
I've spent the last weekend playing against (and beating) the "Incredible" difficulty AI on medium maps. I'm using the patch Brad released on Friday (the first patch to have the auto-pop-from-orbit checkbox).

1. It is my opinion that Cultural palaces need to be removed (or severely nerfed). In my current game, the Altarians were right next to me with more overlords, more tech, more economy, etc, etc. I built 3 starbases with cultural palaces on them in my main system. Combined with my various cultural wonders, that sector was putting out 1.5 MILLION influence points. Needless to say, the Altarians effectively ceased to exist within two years -- a more powerful race than me destroyed for the insignificant cost of 15 constructors.

2. The AI is not intelligent about how it plans its attacks, even on "incredible" difficulty. The AI seems to be content to produce ships and simply fling them at the opponent with no particular strategy. It is slow to follow up with transports, and will often spread its attacks out against a single planet over so many turns that the defending ships have ample time to heal. The AI will ignore starbases unless they are closer than the planets. This leads to the strat of tucking a defense 3 starbase in the far corner of a sector and loading it down with all the +attack modules. The AI will bleed itself white against the super-augmented defenders in the sector without every trying to locate or destroy the starbase that is making the defenders so buff.

3. In general, the strategy of simply sitting back and defending is too easy. Unless the human player is "killed in the cradle", carefully researching your techs and paying the correct bribes will ensure that you can keep your military high enuff to avoid being attacked by the AI. The various AI players will then fight amongst each other, often with massive economic losses (see point #2), enabling the defensive player to get a culture or tech victory with minimal difficulty.

The above is based on my opinions and play experience. It is offered as (hopefully) constructive criticism and feedback. Please don't hurt me. :-)

(the thread in the avault forums dealing with the avault review has me just a teeeeeny tiny bit scared of saying anything bad about the GalCiv AI....)
2,428 views 38 replies
Reply #26 from Stardock Forums Top
What about the diplomacy chees tactic of the day? Wha-huh??

Um... well, if the code was in there at one point, couldn't there be a way (as tugh as it may be) to add some of it back in and let people decide for themselves if it gets used or not. I know I would be willing to wait some minutes for the computer to "think" if it needs too. And I can certainly refrain from Gigantic maps if smallers ones help the AI think.

Arg, I'll give ya more money... -sigh-

I'll stop looking for CHEESE in diplomacy!! That ought to do it, that oughta convince you.

~SDC~
Reply #27 from Stardock Forums Top
I agree with Horatio. We know you still have the code around. Version control, baby. Slap it back in, add an "Advanced AI" setting (heck, it could even be SO-PHIS-TI-CAT-ED and have multiple levels from Celeron 600 Mhz to P4 2.4 Ghz to P4 4.2 Ghz Hyperthread, water-cooled, overclocked mad machine).

:)

/ds
Sins of a Solar Empire II
Reply #28 from Stardock Forums Top
He forgot to say please. So here: p-p-p-p-p-please????

~SDC~
Reply #29 from Stardock Forums Top
Even just checking the closest two systems as possibilities.. and actually checking the defensive abilities conferred to them by starbases.. hmm, example is probably easier to explain.

1. AI checks two closest systems for 'Death by Starbase' (DbS = 2 or 3 fully equipped miliatary starbases)
2. If neither is a DbS system, it attacks one randomly in strength.
3. If one or both are DbS systems, it flags them as such and ignores them - and avoids them with pathing from then on.
4. It then checks the next closest border systems until it has it's two to pick from and attacks one.
5. Perhaps sometimes it deviates and attacks both at once if it's military compared to the target's is strong enough.
6. If the player has ALL their close border systems as DbS systems, the computer rethinks it's war strategy.. and perhaps evaluates it's chances of nuking the closest DbS system outright.

A lot easier here to describe than it is to code it, I'm sure... but that's why Frogboy gets the kudos for a job well done. :p

~SDC~
Reply #30 from Stardock Forums Top
I have to say that the AI that people are outlining would be great.

That being said it sounds very hard to implement. It sounds easy to say, but honestly things that seem easy for humans to do are quite often very complex algorithms.

However he is the list of enhancements proposed that seem to be easy to code:

1) nerf culture. No problem. Heck it would still be easy to nerf culture and have it be an exciting part of the game.
2) Attack starbases which give military bonuses within the escort of current conflict. This does not seem too bad.
3) Bases on resources should give much higher preference to mining. Heck they many want to weigh resource extraction techs more heavily if they have a couple of resources.
4) The AI should consider attacking enemy resource star bases at the start of hostilities very heavily.

Some more difficult ideas would be:

1) The AI should employ more "killer stacks"
2) The AI should consider attacking more than one lightly defended or not defended target per turn if possible.
3) The AI should weight troops ships more heavily (maybe, not really sure about this one)
Reply #31 from Stardock Forums Top
Brad,

I'm not sure why at least a few of these AI enhancements would cause a performance hit. The biggest one that is most desperately needed is for the AI ships to stop simply targetting the closest enemy planet/ship and take into account the overall situation (ie, target the military starbase instead). I honestly think this is a very serious problem, at least for your more serious number-crunching players, and that more and more people are going to notice it and have a problem with it as time goes on. It is just too dang easy to exploit once you understand what is going on. For a game that has a great economic AI and a good diplomatic AI, to have a military AI whose sole strat is limited to "charge at the closest enemy critter" is a bit of a let down (no offense intended).

And I'm not sure why why something like fixing the AI military targetting would scale linearly with the number of ships, as your post implies. After all, the AI should only have to decide on priority targets ONCE per round, and then send its ships towards that target as appropriate. Actually sending the ships at a target is a linear operation of course, but that has to be done anyway. I don't understand why picking a list of priority targets can't be done above the ship level....? Unless if *all* war strategy is being planned out on a per-ship basis, which seems inefficient...? Am I missing something obvious about why this would cause a big performance hit?
Reply #32 from Stardock Forums Top
Hmmm. Just reread Brad's post... I might have misinterpreted what he said, since he didn't say _which_ of the mentioned AI enhancements would scale linearly, while I assumed he _must_ have been talking about the AI enhancement I care about the most. Heh.

In which case, sorry, and never mind my above post (grrr... can't edit...)
Reply #33 from Stardock Forums Top
Unfortunately a AnalyseTheOverallSituation() is not that trivial.

As I mentioned, we used to have more of that in there but one can't just "Go back" to it in an afternoon because so many other things hvae changed since then.

I am commited to making the AI increasingly intelligent. I think everyone here would agree that it is pretty good already.

But bear with me, I have my OWN list of "issues" with the AI that I want to address first. ;) For instance, I want to see the AI handle star bases better.

Ultimately, if the choice is between making the AI more intelligent or combating a cheese tatic, I'll probably go with the more intelligent one. GalCiv is, after all, a single player game. If someone wants to ruin it for themselves, they certainly can do that.

Let me give you an example:
In the OS/2 version, there was a flaw where you could put 3 scouts in a row and block any AI ship. The pathfinding simply couldn't handle complex obstacles. So users would use scouts to block alien colony ships and use that to win every time. But fixing it would have required a quite a bit of effot and made the game slow on the typical machine of the day (hence why pathfinding in strategy games was so poor until recent years).

Coding has to be assigned to things that will benefit the most people. I know of many many exploits in the AI that would be very very hard to work around that are even more serious that what is mentioned here and in fact the reason I had written the overall quaddrant (4 sector) tactical evaluation was because of exploits like you mention. But the performance hit and working out other buglets and issues were not worth the cost in CPU and developre expense when much of what you describe is self inflicted.

That doesn't mean that these things won't get looked at seriously. It just means that they won't have the kind of priority as things that affect more or most players.
Reply #34 from Stardock Forums Top
That all makes sense, Brad.

Bear in mind that perhaps no small number of us would be willing to take performance hits in favor of smarts. After all, This game will be on my HD long after I upgrade my 1800+ it runs on now, and it's not a realtime game- we're already thinkers in this group.

Maybe a frontpage poll to see, like the old SP vs. MP poll that went up months ago.

~SDC~
Reply #35 from Stardock Forums Top
Brad,

Really? AI exploits that I haven't found yet? I'd better get cracking! :)

I don't fully understand the reasons for everything you've said, and I consider the AI combat targetting to be an "unavoidable" exploit... ie, one that you can't help but to exploit, simply by the very nature of wanting to most heavily defend the places that get attacked. I mean, sure, I won't ever use the Cultural Palace again, nor will I ever build cultural improvements except defensively. But I can't choose to attack stupidly, or not defend my planets, or something like that.

But then again, I can't see the code and you can, and it isn't like you are under any obligation to explain yourself to every problem child in the forums. :) And as I'm sure you know, you are going above and beyond the call of duty in supporting this game. So having made my concerns clear, I shall now back off and say nothing further about it, at least on this particular topic.

Thanks for the great game and happy coding! :)
Reply #36 from Stardock Forums Top
I wouldn't be too quick to toss out performance in favor of better AI. I mean, I like having a good, challenging AI that *does not cheat*, but I don't want that at the expense of being able to run the game on my system.



~SDC~
Reply #37 from Stardock Forums Top
Well I think you should update the gameplay examples to show you using all the AI exploits you know and winning the game comfortably. ;)

The way you make it sound there, it's like "The AI I designed is so brilliant even I can't beat it", where as the impression I'm getting from the forums is that this isn't really an accurate portrayal. ;)
Reply #38 from Stardock Forums Top
Iskabis, your analysis is correct (I did not have any ships in any of my planet, so it just targeted the starbases). And I can really understand what Brad is saying about the game slowing down too much if each ship starts to do too detailed calculations (I remember some early betas being really slow when lot of ships were around and it speeded up a lot when Brad changed the ship handling AI).

Brad - what are you meaning with "make the AI even more intelligent in basic strategic terms"?

Does this refer to overall startegy or the ship handling strategy? Improvements in overall strategy can be probably done still quite easily without adding too much calculation. But I cannot easily find ways to improve the ship handling without adding much more calculations (ok, maybe better stack building by using defenders, attackers and transports in one stack and having it near the planet before announcing the war).

Periz