It's just sad that the AI is not so good

[zulu9812]

The AI in Civ3 is much, much better than in previous games. It's predictable, but it puts up more of a fight.

[Capt Dizle]

Darn if I can see it. Maybe you are mistaking AI production bonuses and the movement restrictions for AI cleverness.

BTW, I am no longer seeing the AI self induced starvation on a mass scale. Are you guys still seeing it?

[zulu9812]

Perhaps. At higher difficulty levels, it'd be nice to see the AI employ new strategies rather than simply gain handicaps.

[Arrian]

BTW, I am no longer seeing the AI self induced starvation on a mass scale. Are you guys still seeing it?

They lowered the poprush/drafting penaty back to 20 turns from 40. That helped prevent them from getting so unhappy that they starved their own cities. The AI *may* be using those things a bit less, but I can't be sure.

-Arrian

[Panag]

Re: It's just sad that the AI is not so good

Originally posted by jimmytrick
Despite the bragging the AI is just not up to par. I find it kinda sad. I know I'll be flamed to **** and gone but it doesn't change the facts.

A simple barrier of infantry, two to a tile, will cause an invasion force of a hundred or more units to turn and sulk home. Invade with a transport full of armor and despite the fact that the AI has a full RR net and 250 units, it will only defend a city with a handful (3-4) of units. The AI has no ability to support attacks with air and artillery support. Heck, it can't even support a defensive position.

Firaxis turned Civ into a wargame and claimed the AI was great. But it's not.

The only real impediment to military conquest is the inability to use enemy road and rail nets.

I guess some of you are easily satisfied or something.

hi ,

put the default level for the AI , in the editor at deity , ...

have a nice day

[Fitz]

I think the real problem is teaching the AI to differentiate between a strong threat and a weak threat. Threat assesment that is relatively simple to a human can be extremely difficult to program into an AI.

Example: Human has 3 workers in place next to 5 modern tanks. The obvious lure to a human is to see that the workers are intended to move you onto less defenesible terrain. However, if you just reduce the AI's tendency to take workers, this means that they will sometimes fall for the lure (still), and othertimes will ignore workers it is perfectly safe to take.

Example2: Surrounding your borders with infantry or other defenders makes you impregnible to attack. Now, the AI makes very little assessment of its ability to break through this old fasioned defense. I have seen an stack of cavalry and infantry take out a single border guard, but when you block of their next advance, they retreat.
Admitedly, example two may be a good example of AI successfully realizing that you have massive defense in depth, and troops ratio comparisons. But instead of a way around it (Air & Artillery bombardment, attempting multiple breaches), it just gives up the ghost.

Jt's example of the transport is another example of insufficient threat analysis ability.

Add to the poor threat analysis bad tactics, and you get an AI that cannot defend itself effectively and cannot attaack sufficiently, and is only successful so long as it has a (frequently significant) lead in number and firepower of troops. Whereas its human opponent can easily fend of the AI with less troops of inferior quality, and decimates it when it is equal.

[SultanofATL]

I haven't seen the AI quirks mentioned by JT. What I've seen is quite the opposite. I've had stacks of my best Units stacked in hidden defensive positions waiting to pounce on the enemy should it invade. The computer on several occasions has entered my territory allowing it to see this stack and then launched an all our assault on the position. The only reasoning I can give to this behavior is the acknowledgement of a significant threat. I've also had galleys and troop ships intercepted and followed by friendly AI once I entered their territory. Again I conclude this to be an escort to make sure I don't sneak attack them. I don't have a save game to show these particular instances, but now I will try to save these instances for proof. Maybe its just my version?

[Panag]

Originally posted by SultanofATL
I haven't seen the AI quirks mentioned by JT. What I've seen is quite the opposite. I've had stacks of my best Units stacked in hidden defensive positions waiting to pounce on the enemy should it invade. The computer on several occasions has entered my territory allowing it to see this stack and then launched an all our assault on the position. The only reasoning I can give to this behavior is the acknowledgement of a significant threat. I've also had galleys and troop ships intercepted and followed by friendly AI once I entered their territory. Again I conclude this to be an escort to make sure I don't sneak attack them. I don't have a save game to show these particular instances, but now I will try to save these instances for proof. Maybe its just my version?

hi ,

nope , others have seen it to , ...

have a nice day

[ChaotikVisions]

The AI does utilize its RR network for defense, i've had smaller landings of 8-16 Tanks completely wiped out by the AI on the same turn. And they also do use Air power, but it seems they only want to target ships and regular units out int he open, not citys. Also, the AI seems willing to lose smaller citys for some reason.

[Dinner]

Originally posted by Fitz
As an example, if they took just a few of 'Polys best players, watched them play a few games, and started incorporating their strats into the hardest level, you would have a game that was designed for the public to play on standard levels (and not give up in frustration)

The sad truth is Firaxis has already confirmed that the up coming patch as well as PtW will not do anything to improve the AI. It looks like MP is the only avenue left for those of us who would like a more intelligent opponent.

If Firaxis had included a scripting language then we could work on improving Civ3's mediocore AI just as people here at Apolyton improved CTP2's AI, however, it doesn't look like Firaxis will be doing that either...

[Panag]

Originally posted by ChaotikVisions
The AI does utilize its RR network for defense, i've had smaller landings of 8-16 Tanks completely wiped out by the AI on the same turn. And they also do use Air power, but it seems they only want to target ships and regular units out int he open, not citys. Also, the AI seems willing to lose smaller citys for some reason.

hi ,

the AI also try's to minimize its loses , .... and sometimes regroups , and if for one he has to give a city up , he shall do so , sometimes he sell's everything there is in the city , ....before he lets it go , ...

have a nice day

[DeepO]

This is a welcome change from previous AI complaints, as here a lot of well defined points of possible improvements are given. So I'll add my comments to the mix

First of all, I do find this AI remarkably better than previous games. Other games may have more focused AI, but you have to remember that Civ is very complex, it is very difficult to build an AI that's good in all aspects.
The suggestion that Firaxis should look at how vet players are winning might be a little late, I'm quite positive that many of the added AI routines resulted from looking at these boards.

Originally posted by jimmytrick
1. AI will invaribly advance through gaps in defensive positions.
AI will often retreat when such gaps are closed. AI will not attack strongpoints.

Solution: Change the pathfinding so that the AI will be able to differentiate between impassible terrain and troops that can be killed. Add variable routines so that the AI will sometimes attack strongpoints rather than bypass or retreat.

True. However, in certain cases strongpoints get attacked, it's just not very strategic in deciding which ones are attacked, and which are left alone. It seems to me that, when an army would have advanced at a particular place without the strongpoint, it will also attack the strongpoint. This should be addressed. period.

2. AI does not use artillery and air support to soften strongpoints.

Solution: Instruct the AI to build and use artillery in mass. Instruct the AI to build airfields and support offensives with air support.

This also depends on the situation, I've seen the AI softening up the terrain with bombardments before it attacks. However, strongpoints & cities should be given a higher importance to get bombed. As it is, most artillery seems to be used as a passive defense, in that many cities will produce artillery, but keep it inside the walls to counter frontal attacks with free fire. If a unit gets in range it will be bombed, but only if that means that the passive defence stays in its place. This could be improved, artillery is foremost an offensive unit, or a actively defensive one.

3. AI does not utilize RR system properly for defense. AI defends passively and does not use artillery or air support for defense.

Solution: AI should be taught to aggressively attack invasion forces (leaving a random chance of delayed or passive response to prevent yo-yoing.)

I agree that the RR system is underused by AIs, but not like you mention. One of my latest games I had a fleet of battleships bombing strategic resources of an AI. Whenever a BS stopped in range of an artillery, it was fired upon. However, as the BSs moved along he coast, the artillery stayed behind in the original city, instead of following it on the RR, and bombing again from a different city. Should be addressed.

Other than moving artillery, I think the AI will use RR quite well to counter you, if you go in with small strikeforces to deny the AI some resources (so no ultimate blitzkrieg with superior force), units tend to come from all over the place, and are more or less used to their best effect.

But, the order in which units move is strange: a human player will first counter an attack force, and then divide the remainder of his forces to get in a good position for next turn. The AI seems to do both at the same time: an army is attacked by several units, while in the mean time other units go right passed it to some other place. Sometimes, due to the randomness of combat, the AI does not have enough forces to totally anihilate your army, and those last 2 units stay behind with one HP or so. This is nonsense, as there were so many units passing by, which could have killed the unit and still get to where they wanted to go with the rest of their movement points.

4. AI grabs workers.

Solution: Reduce AI tendancy to grab workers.

Solution is not to reduce the tendancy completely, but to have an assessment of how valuable that worker is, as opposed to the risk the offensive unit is taking when grabbing it. In certain situations, grabbing workers is vital, in others it doesn't matter as there are so many running around. The AI is over simplified in this respect, but in order to prevent it, it could be needed to give the AI full troop inspection possibilities of a human; an unfair advantage for many including myself.

5. AI does not recognize threats.

Solution: While impossible to teach the AI to "see" threats the way humans do, it should be possible to flag certain things and link up AI reaction. Example: A fully loaded transport within one turn of landing troops should provoke some kind of reaction from the AI if it has been spotted. Can surely teach the AI to line coast with units to require marines (which would give the marines a purpose in the SP game).

I agree more or less. It will be very hard to program a good threat algorithm what ships are concerned, certainly if you want something that also works before war is actually declared (standard mass landings happen in the same turn that war is declared, or you haven't planned it right). As you can't see how many units a transport is carrying, you can ony have a guess at it's intentions. Is that transport going right passed your coast, or does it plan on landing here? Are there any units in them, or does he go to load them up? Too many variables, I'm afraid.

And of course you could 'teach' the AI to line it's coast with units, but I feel this is applicable in certain situations, not all. Some algorithm could be devised wich says that if you have 20 times more units then coast tiles, you put one of them on each coast. Or something like that. But don't waste units on lining coasts, when you have in land choke points lowly defended.

6. The AI is predictable.

Solution: Randomize. Even stupid responses can be effective, but predictable responses seldom are. Don't force the human to play poorly to simulate meaningful conflict.

This is something I totally agree with. However, randomisation is quite risky while keeping a stable AI, and should be done with extreme care. In the path finding, for instance, there should be some factor added which makes the equivalent paths (in terms of turns to reach a target) all possible for a unit, instead of taking allways the same path. This comes from the algorithm which will take the first solution of the A* routine: it will first try the east direction, cycling clockwise (or south and counterclockwise, I haven't tested the actual parameters), and will always use the first, best direction. While it is garantueed that the path taken is the shortest one, it is always the same, which could be easily spotted by a human player, and anticipated with some strongpoint along the way.

Randomizing this direction (starting from a random direction before cycling) does give better results, even if it is computationally a bit heavier. However, you'll need additional guard algorithms for stack movement. Right now, when a settler and its defender want to go together to a certain spot, you first direct the settler, then the defender towards that spot, and the identical path they'll follow will ensure that at each turn the defender is in the same tile as the settler. Once you have randomized path finding, you're not sure of that anymore, and you'll need more micromanagement to make sure the defender is following the settler, instead of taking another (shortest) path to the destination.

The same is true for armies, my guess is that they all use the same path finding routine, over and over again, instead of have some kind of stack movement routine. This is not ideal, but changing it would require a totally new view, which is no longer a simple adaptation of existing software.

7. The AI does not upgrade.

Solution: Upgrade. This should be first on the list and is the most unforgivable.

In 1.21f this is simply no longer true. When the AI has the spare money, it will upgrade everything it can. I find myself devising tactics that will ensure that an AI does not have spare money, just to prevent them to attack me with massive stacks of modern units the moment they get a new tech.

What might be improved is that it seems that the AI does not save any money on purpose: it will instantly use money for whatever reason it thinks best, instead of having some put aside for upgrading because it can expect to have an upgrade available in one turn.

Originally posted by Fitz
I think the real problem is teaching the AI to differentiate between a strong threat and a weak threat. Threat assesment that is relatively simple to a human can be extremely difficult to program into an AI.

Completely true. Threat assessment is very hard for a rule based system, it requires some kind of patern recognition that is intuitive to humans, and can't be easily transformed in fixed rules. Some simple things could be added, but than chances are that those rules are as easy to counter as without rules now.


One of the few things I would love to see improved is that the defense of AI cities should be adapted to the risk of being taken. Now there simply is a standard '2 defender per city' rule, with more troops added when they are available. Sometimes you see 3 defenders in very important cities, very rarely 4 (this could be coinicidence, as those cities in many cases are the most productive unit builders). There should be an improvement in that very important cities (resources, wonders, capital) should be defended with more units. If I have an army of 80 MAs advancing to the capital, you would expect at least 10 infantry to hold position, not 2.

One other point is that a SS should also be a goal for the AI, instead of something that is achieved by accident when every tech is available. It sure feels like the SS was included to give every human, on all difficulty levels, the chance of winning the game, even in an underdog situation. It should be a real goal for those civs that do not have the chance of winning military, not just an accidental thing.

Final remark to Coracle: it is very obvious that you don't know sh!t when it comes to AI. If you think a learning system is easily programmed, think again. The only modern game which had a learning AI was Black&White, and after it came out many people (including myself) were disappointed in the limited amount of learning the AI could put up with. And this is in a game that was build around a learning AI, a decision tree with believe system. They had three software engineers working on it for several years, and still it was disappointing.

I recently started a PhD on learning, adaptable AI (mostly in agents for knowledge management), and even in a protected lab environment it is extremely hard to have something which performs well. There are simply too many parameters involved, and even if games are at the summit of technical feasible techniques, you cannot expect to have learning in a game with so many complex sides as Civ. It would be very cool to have a TBS that could learn from your tactics, and devises tactics of it's own to counter the human, but we are at least 10 years away for that. Maybe in Civ VI.

DeepO

[edited to fix vB tags]

[Capt Dizle]

DeepO,

don't have time to respond as I would like, just two quick points. In a v1.21, huge, 16 civ game, the English are agressively going for a space victory. This includes agressive tech research and immediate component building. This same civ refuses to use RRs to counter my attacks.

I swear I can't see any evidence the AI is upgrading much. More later.

[DeepO]

The English always had a mind of their own
I never saw aggresive SS from an AI, while I haven't found many games where the AI refused to use RR. But this only shows that we should be careful when it comes to interpreting our own experiences as game facts.

DeepO

[Panag]

hi ,

intresting DeepO , ....

have a nice day