I like adding a bonus to AI for difficulty. I wonder if 'personalities' will be from race to race, or if each race will have more than one personality.

Becks wrote:
Very nice to read this. But I have one question. Is the AI able to learn from the behavior of the player?

As awesome as a self learning a.i would be. I imagine it's very hard to programme though otherwise i think more games would have used it. I could be wrong of course as i know nothing about programming.

The hardest with learning AI is to... learn. You need a lot of iteration to get some not so bad AI. And they need to battle not only each others but also humans to be at least effective.

VieuxChat wrote:
The hardest with learning AI is to... learn. You need a lot of iteration to get some not so bad AI. And they need to battle not only each others but also humans to be at least effective.

The typical problem for that kind of thing is "overtraining". For a program to learn, you set a range of input variables and define your desired output, e.g. winning the game. Then you have to put it into the environment it will be used in, the game e.g., let it work against s.o. or s.th. and then repeat it until you've found a set of functions of your input variables that work best at giving you the desired output. That is called training.

If consequently done against s.th. that is rather stable, like another AI with fixed parameters, this method will of course win in the long run, but might totally fail against a human who adapts.

If you train too long, you might end up with an AI that is capable of only winning that one scenario you trained it for and totally unable to do anything else.

If you don't train it long enough, it might be a bit of a loon and do totally unproductive things.



And making it learn during one game in itself, that is really hard to achieve. You'd probably have to do s.th. black and white-like and repeat some feedback on some output of the AI constantly for it to have some idea where to shift its priorities. In a 4x-game that might just as well mean the faction the AI controls loses before it achieves any kind of normal gameplay chance.



Correct me if I'm wrong on any of these points, please. There should be a lot of people in this forum who know more about AI controlling than I do of multivariate analyses and programming. XD

I don´t think there´s ever been a relatively complex game in which the AI doesn´t cheat, specially strategy games. It´s good enough if the cheats does not get noticed.

Damntry wrote:
I don´t think there´s ever been a relatively complex game in which the AI doesn´t cheat, specially strategy games. It´s good enough if the cheats does not get noticed.

I have to agree. Since they aren't players, they're the game itself.

Becks wrote:
If I remember right, SpaceTroll said that "you can compare the AI to a baby. A complex brain but it doesn't know how to use it well." Therefore I think that the AI will not cheat. The baby needs only some training.

How is babe formed... :eek:

Well if its a baby, then it must have really caring parents cause i find myself constantly encountering 2 map size civilizations by the time i colonize my 6th planet (~Turn 50)

In my mind, giving a "bonus" to AI FIDS output is cheating.

VieuxChat wrote:
The hardest with learning AI is to... learn. You need a lot of iteration to get some not so bad AI. And they need to battle not only each others but also humans to be at least effective.

This can be done "genetically" with a reporting function back to amplitude. treat every game as a data point and go from there. Slowly, the AI would find a strategy that works against most comers. You can force the AI to balance against each other too by running playerless games; This would probably be the starting basis for a genetic AI. Anyway, They are looking to sell a 4x, not make breathtaking AI breakthroughs. Its probably quite a hard problem to tackle.



I am curious about the agent in-game learning. If I build a laser heavy armada, do they move to counter my change in tactics quickly? That sort of thing. Round to round learning.

Very nice to read this. But I have one question. Is the AI able to learn from the behavior of the player?

As there are occasional questions on the forums about the AI and how it works, I am giving the thread a little nudge.

Then I'm going to do my best keeping it up for a while, since I have to add something myself anyway.



I think it's fine if the AI cheats a little, it just has to be done in a specific way. Increasing the FIDS output on the AI is one way of cheating I deem fine. I remember Space Empires 5 giving you the option to let the AI cheat separately from all other options and I always used it to get a little more challenge. As for ES I think the AI does kind of okay during the game but it has to become somewhat stronger, I think the AI is good when on impossible the player has to be at least pretty skilled to keep up with them. So far I'm plenty able to stomp 2 empires on Impossible. Without any wars between any of the factions I'm usually about 1,5-2 as strong as both enemy empires at round 100 on fast, both enemy empires together that is.



But appart from the AI still being a little weak I really like what I see and after reading ThorTillas post I'm optimistic the AI is going to be very good upon release.

Good thread I'm wondering if the AI factions will work together when faced with a common enemy and if they will be able to incorporate strategy derived by players and posted on these forums (examples. Destroyer swarms, Immediately attack the Cravers, harassment wars to disrupt enemy FIDS etc.)

stevenmc409 wrote:
Good thread I'm wondering if the AI factions will work together when faced with a common enemy

That's a tricky one. Alliances between factions to beat another (non-player) faction could be interesting. The typical "gang up on the human player when he gets too big" alliance is something else though. It's become a convention in strategy games to slow down a winning player, but it's annoying when it's too predictable.

and if they will be able to incorporate strategy derived by players and posted on these forums (examples. Destroyer swarms, Immediately attack the Cravers, harassment wars to disrupt enemy FIDS etc.)

IIRC, the way this was done in GalCiv2 was that the devs would track player strategies discussed in the forums, and then modify the AI routines to cope with common player strategies as each update was released. That's a more reasonable goal, obviously, than expecting the AI to learn on the fly. If the game is continually updated after release, I hope that's the plan.

I played my first game on easy and the AI was outproducing me three to one by the time I saw it for the first time, and two to one right from the start. He showed up at turn eighty with significantly more, better, larger, and bigger fleet stack. Except for the wormhole bottle neck and some early Hero wins the game would have been done. It had a massive advantage, a seemingly endless supply of ships, and also had the map covered well before I did my half. Now this was my first game and all but there is no way things were equal. I never lost a battle for the first 140 turns, and hardly lost a ship BUT the AI was still ahead on points 3 to 1 by the time the game died to a heap error. So after blowing up hundreds of ships, maxing one Hero out at 20 and the others almost maxed out as well and it sure felt like the AI still had 10-15 times more ships then I did.



I'll give it a few more tries but my first impression is the AI needs some significant tweaking on the easy level. Not to be Captain Obvious or Melvin McCranky Pants but easy should be... well easy. Its where you learn the ins and outs of the game play.



It just felt like a tremendous grind for each planet, endless enemy spawn, and never enough ships to plug the gaps no matter how many times I beat the incoming fleets



Oh and normal should be a challenge for the average player.

Nosferatiel wrote:
*tinkers a bit the summon dev spellbook, black candles, badly drawn pentagrams included, blood sacrifices and gore excluded*

I think the AI can cheat a little bit, if not it just has a better area to start out, but most of all... I must say I'm quoting that if I have your permission Nosferatiel (ex: Facebook and the likes)

Some of the AI behaviour seems a bit suspicious to me.

Such as the way when, at first contact or immediately after declaring war, I probe an AI race with one or two small ships using kinetics, to see what they are using (i.e. mainly missile with a little kinetic), I send in a couple of large heavy fleets with plenty of flack & deflect armour with invasion mod and beam weapons. All is fine and I wipe out the defense fleet so fast that only 1 or 2 ships get their missiles off and I take almost no damage at all and have 2 full battle rounds to heal up. Very next turn, the enemy sends in 10 full fleets from an incredible distance away and the damage my laser weapons do drops from 50+% to less than 20% and both my fleets are wiped out. I mean, that AI was real quick to suddenly research shield defence, get massive movement bonus, have the dust to upgrade all those fleets and not miss a single turn of movement in getting them into my face.

I think the AI already cheats.

isaristh wrote:
I'm going to hope the AI doesn't cheat, or at the very least an option is given to enable or disable cheats. Quite a few games in the past have really ticked me off when the AI gets handicaps and advantages, sometimes when they don't really need them. I can understand cheats on harder difficulties or something, but there needs to be an option where the AI is no better off than you are.

They're called restraints - designing a thinking AI is not something you just do. Especially not in games as complex as these. Sorry but you're going to have to live with some cheating.

must say this makes me chuckle



Originally Posted by Nosferatiel

I'm just curious if techniques I've seen used in physics actually have interesting other properties than searching for, let's say, the goddamn Higgs particle. XD



so i had to add this makes youre mind allmost exsplode

http://io9.com/5905897/the-webcomic-guide-to-the-higgs-boson

metaspirit wrote:
just as the title says will the ai cheat i blew the dust off moo2 and it struck me.

Just how much the ai cheats:eek: you know like 5-10 turns into a game and you meet another race and they already have 5 to 8 worlds. Its like they have x3 speed its just like fighting up hill so whats the ai going to be like ? anyone know

If I remember right, SpaceTroll said that "you can compare the AI to a baby. A complex brain but it doesn't know how to use it well." Therefore I think that the AI will not cheat. The baby needs only some training.