Rational AI please?

This thread is silly. 2 points

1) I have many complaints about the AI. Nothing in this post was a valid complaint. In fact, they played how I would expect them to. Or did you perhaps want him to just roll over while you win? In multi-player, if my opponent is rapidly expanding I can tell you what they aren't building...units. I would see this as an opportunity and do the same thing as Denmark. This is GOOD AI.

2) You obviously have no idea about AI and the limitations. People learn, computers do not. Everyone always mentions the computer that beat the world champion at Chess. Well it did, but guess who won the 2nd and 3rd time? Also, like others have aid, Civ is an incredibly complicated game to program for, there are so many variables.

3) Yes I know I said 2 but this is my third. CnC? Seriously? I play that game as well and the AI is NOT better than civ. First of all its a lot easier to program for. Its just a real time war game. Strategy: Attack with effective units. Destroy opponent. The AI cheats horribly on the hard and brutal levels in that game as well...its almost more ridiculous than Civ since its a real time strategy and the computer already has an advantage in being able to control multiple units at once. I remember playing CnC 4 and tried it on brutal. The AI rushed me with 10 tanks in what couldn't have been more than 2 minutes. The player can't physically defend against that so early. I had 1 tank at that point. That wasn't good AI, that was giving the AI unlimited funds and severely reduced build times. Neither of those things are "strategic".

Yzman said:

This thread is silly. 2 points

1) I have many complaints about the AI. Nothing in this post was a valid complaint. In fact, they played how I would expect them to. Or did you perhaps want him to just roll over while you win?

Click to expand...

Well, it's Harald so that would be normal...

Dogmouth said:

Command & Conquer is a real time strategy game, so it's not comparable. Real time AIs have a natural advantage over humans because of limitations in UI (inability to enter multiple commands at once) and limitations in the breadth of human processing (our multitasking abilities are much worse than a garden variety computer). Real time AIs are not comparable to turn-based AIs.

Click to expand...

And even given this they can still produce comparably bad results. Strictly speaking, something like the Starcraft II AI is worse than Civ V AI - not because it has fewer tasks to handle, but because it has very little adaptability. It amounts to an automated build order, attacks consistently at time X with little allowance made for what the opponent has, and it can wipe out your army and then run away because 'run away at time Y' is a good general rule but it lacks the ability to identify that your base is undefended and adapt its play to press the attack. Civ V AI makes a lot of bad decisions, but it is actually sophisticated enough to make decisions rather than to play by rote.

Yzman said:

2) You obviously have no idea about AI and the limitations. People learn, computers do not. Everyone always mentions the computer that beat the world champion at Chess. Well it did, but guess who won the 2nd and 3rd time? Also, like others have aid, Civ is an incredibly complicated game to program for, there are so many variables.

Click to expand...

You mean games or matches, because there was only one match with that computer (and one with an earlier IBM computer), and the computer won. If you mean games, Kasparov did better at first but lost ground as the games went on. If you count the first computer too, that computer lost, and the 2nd one won.

Computers are able to "learn." We're likely just a ways off of allowing your home computer to handle the AI to "learn" how to play Civ, considering Pong is hard enough for now.

Rooftrellen said:

You mean games or matches, because there was only one match with that computer (and one with an earlier IBM computer), and the computer won. If you mean games, Kasparov did better at first but lost ground as the games went on. If you count the first computer too, that computer lost, and the 2nd one won.

Computers are able to "learn."

Click to expand...

They were two different programs, created several years apart. Deep Blue didn't 'learn' from Deep Thought's mistakes.

Certainly neural nets have been developed - indeed one was used by MIT to learn how to play Civ V, and reportedly won against standard Civ V AIs. I also hear it made turn times excessive, and beyond that I don't think it's been specified just how powerful the machines designed to play it were.

We're likely just a ways off of allowing your home computer to handle the AI to "learn" how to play Civ, considering Pong is hard enough for now.

Click to expand...

This is probably true in much the same way that it's been "true" for 40 years that human-level AIs are 20 years in the future, or that commercial fusion power is 50 years in the future.

PhilBowles said:

They were two different programs, created several years apart. Deep Blue didn't 'learn' from Deep Thought's mistakes.

Certainly neural nets have been developed - indeed one was used by MIT to learn how to play Civ V, and reportedly won against standard Civ V AIs. I also hear it made turn times excessive, and beyond that I don't think it's been specified just how powerful the machines designed to play it were.

Click to expand...

Deep Blue was modified from game to game by IBM to cover up holes in it's "strategy." It wasn't what I was referring to with computers that can learn. IBM created a chip that could "learn" how to play Pong, and I read some interesting stuff on how computers make optimal designs for things such as antennae though trial and error, and can use what they "learn" from previous designs to do better on their future attempts.

I'd be scared to imagine what my computer would do if it tried to run an AI that made turn times "excessive" on a (or several) computer at MIT, but that's why we need to give the AI bonuses to make it harder. Heck, even THAT probably couldn't give a human much challenge without any bonuses at all.

smallfish said:

Well, they certainly know how to goad you into attacking them. Make of this what you will

Spoiler :

Click to expand...

That'll definitely work on me if it happened in a game of mine unless I somehow don't care about him xD Thank u for the share.

And well, AI is plenty rational in civ 5 compared in civ 4 where i could just sleep my way through the game while having metropolislust by building as much huge cities as possible. i used to always try to see if city graphics would expand and take all 9 tiles or not, it only grew to the point of taking half of the tiles next to the center tile. And bazillion stack of dooms on defense.. <.< The AI really couldn't threaten me in civ 4 compared to civ 5.

The ai will be happy to kill you off if you're weak unless he or she REALLY likes you in civ 5.

In one of my older games from last month or 2, I went for world domination but, sweden and netherlands was still friendly and cool with me, I gave up on world domination and even wasted 5 turns of total war buff trying to push myself into killing netherlands off first then sweden and gave up and declared the world at peace lol with all the foul mouthed ais dead except for two buddies i have xD And, both netherland/sweden is pretty large as well. They was pretty respectful despite going order with me going autocracy. To go for the kill would leave me with a bad taste in the mouth unless they chose to attack me xD And netherlands even saw me mass the troops on his borders and back and he didn't let a insult fly into my general direction.

That was smart of him, he survived because I couldn't bring myself to go and kill him off. Sweden too. Then I retreated into period of self building, except, problem is, I reached the end of tech tree so i jus tshut the game down.

I love this debate. I'm a CS student and a PHD (two weeks in!) in statistical learning, I know how utterly hard it is to make a computer run such a complex task.

There is something called the combinatorial explosion used to describe complexity of solving such task (civ is obviously EXPTIME-complete). If you take a game with only few simple rules like chess, there is 10 millions different positions 7 moves in, and 10^30 combinations of moves to explore If you want to plan 7 moves ahead (as best players do).

As for civ, where you should be playing more or less 300 moves forward, and take into account a bunch of violently non linear rules, so there is not enough computational power in the universe to solve such an optimization. That's why you can't really compare deep blue (brute force more or less) and civ AI.

In those case the only way is to rely on heuristics, little tricks or patterns that you add to your program to help it narrowing down the research of optimal moves, for example if you have military superiority and proximity and territorial inferiority then Dow. I think the dev did a very good job on diplo and empire building in GnK.

The combat AI is weak because combat is very terrain dependent so it is hard as hell to find heuristics that work every time. But it is not entirely impossible and as patches go I'm expecting to see the worst AI move (great Marchand or space ship part attacks!!!) go, and better fight patterns emerge. When the modders will have the c++ we also might have something good emerging.

PhilBowles said:

They were two different programs, created several years apart. Deep Blue didn't 'learn' from Deep Thought's mistakes.

Certainly neural nets have been developed - indeed one was used by MIT to learn how to play Civ V, and reportedly won against standard Civ V AIs. I also hear it made turn times excessive, and beyond that I don't think it's been specified just how powerful the machines designed to play it were.



This is probably true in much the same way that it's been "true" for 40 years that human-level AIs are 20 years in the future, or that commercial fusion power is 50 years in the future.

Click to expand...

I have to check on that but I don t think it was so much a neural net as memory based learner trained .derailing the topic ?

eidt : oh you were right , neural network it was
We employ a multilayer
neural network where the hidden layers represent
sentence relevance and predicate parsing decisions.
Despite the added complexity, all the parameters
of this non-linear model can be effectively
learned via Monte-Carlo simulations.

Hmm. I just got GK and was fighting the Japanese on Prince. They were actually quite good at first, they got angry at me, but didn`t attack because an African Civ was in the way. they went to war with that CIV and destroyed it, then attacked me.

I fought them back and reduced him from his 7 cities to 2 (his capital and a tiny city). He finally agreed to peace. I decided to let him live. All good.

Next turn he DENOUNCES me. What the? I`ve just reduced him to almost nothing, I haven`t even removed my forces who are still sitting close by and he denounces me? A wise man would`ve kept quiet, be glad his civ was still alive and waited a good few turns until I seemed to have moved my forces away before even looking at me sternly. I could have so easily just moved my forces in and taken his capital, just like that.

Socratatus said:

Hmm. I just got GK and was fighting the Japanese on Prince. They were actually quite good at first, they got angry at me, but didn`t attack because an African Civ was in the way. they went to war with that CIV and destroyed it, then attacked me.

I fought them back and reduced him from his 7 cities to 2 (his capital and a tiny city). He finally agreed to peace. I decided to let him live. All good.

Next turn he DENOUNCES me. What the? I`ve just reduced him to almost nothing, I haven`t even removed my forces who are still sitting close by and he denounces me? A wise man would`ve kept quiet, be glad his civ was still alive and waited a good few turns until I seemed to have moved my forces away before even looking at me sternly. I could have so easily just moved my forces in and taken his capital, just like that.

Click to expand...

No? He's basically lost the game. His best bet is to try to make your situation as bad as possible in hopes people will declare war on you, then make a desperate final assault on you. He's literally a man with nothing left to lose.

Rational AI to me means a Predictable AI which is bad, imo. It also means a dumbed-down (or easier) game where the human player can dictate the tempo and build order without serious opposition. Civ is not a city-builder simulation but a game against opponents that need to be much more aggressive and better at winning (i.e., beating the humna player).

Peng Qi said:

No? He's basically lost the game. His best bet is to try to make your situation as bad as possible in hopes people will declare war on you, then make a desperate final assault on you. He's literally a man with nothing left to lose.

Click to expand...

In fairness, I think this is something that does need to be looked at. It does indeed make sense for an AI, but (a) the AI has difficulty judging context with denouncements (it will often make them on a duel map, which is pointless, and can often use them when they're the civ everyone hates - which is counterproductive because they then get the 'you denounced our friend' negative), and more importantly (b) you know that if you declare peace you will invariably be denounced next turn, which means getting a new negative modifier. This means that it's never in your interest to declare peace if you're winning a war - even if you don't actively attack the other civ, simply remaining in a state of war prevents them denouncing you.

Civ is not a city-builder simulation but a game against opponents that need to be much more aggressive and better at winning (i.e., beating the humna player).

Click to expand...

No, the AI definitely needs to be rational to present a challenge (no more building spaceship parts and not moving them to the capital). It needs to be better at winning, but as no AI is ever going to beat a capable human in combat, the more aggressive the AI, usually the lower its chances of surviving and winning the game. Civ V's AI is already ahead of older Civ games in being able to beat a human to science or cultural victory on occasion, but needs a lot of tightening up, but at domination it's universally acknowledged to be a lot worse than its predecessors.

The biggest thing about all this is I don't understand why they even have modifiers or relationship statuses when it can change so drastically. I've had it both ways. Siam was DoF for hundreds of years and suddenly DOW! And I've also had someone Hostile for hundreds of years, and I give them one luxury for free and DOF! They need to make long time extreme relations take more time to change.

PhilBowles said:

In fairness, I think this is something that does need to be looked at. It does indeed make sense for an AI, but (a) the AI has difficulty judging context with denouncements (it will often make them on a duel map, which is pointless, and can often use them when they're the civ everyone hates - which is counterproductive because they then get the 'you denounced our friend' negative), and more importantly (b) you know that if you declare peace you will invariably be denounced next turn, which means getting a new negative modifier. This means that it's never in your interest to declare peace if you're winning a war - even if you don't actively attack the other civ, simply remaining in a state of war prevents them denouncing you.

Click to expand...

I've actually made DOFs with AIs with whom I have been at war.

AI uses nukes when you nuke them. Sounds rational to me.

I just had Germany pull off a nuclear blitzkrieg on me, taking back one of their cities I captured with a combined arms assault. Just when I thought I had one of my submarines safely tucked in an unseen corner of his border, shooting down stragglers... bam, half of the Kriegsmarine swarms out of the fog and kills it. My only puzzlement was why, despite being at war with America and China on other continents, Bismarck did not bother sending them to knock out more cities for the Reich and instead tucked all of them units, tile-to-tile, around the Fatherland.

As an aside, the new X-COM previews had me going back to play UFO for the first time in years. A very different game, to be sure, but I was struck by the observation that the AI for this 1993 game seems to be at least as good as that in most modern games. There are typical slips - the aliens routinely, and repeatedly, mind-controlling agents who've dropped their weapons, their tendency to stand around uselessly in ships - but they'll set up in strong positions, make good use of cover, retreat if spotted, remember where they last spotted an enemy and move accordingly, and make good use of opportunity fire. Strategically, they'll concentrate later-game attacks on territories they learn you struggle to defend, and actively attack your bases, while UFOs recognise when they're at an advantage or disadvantage and attack/retreat accordingly.

The QC said:

"Doesn't have any troops" counts as provocation . Ignoring the military side does not work in Civ V, you need to have enough troops to defend yourself, at least. Try to go slower with the settlers and to build more units in between.

Click to expand...

This. You were rapidly expanding (ding 1), settled too close to them (ding 2- and another civ won't always say anything to you about it, even though they will ding you for it), and you were weak (ding 3). The Danes are the vikings, after all- agressive attacks on weaker opponents is their forte.

And you need, need, NEED to make some archers, to protect your early cities. Plain warriors suck, and are the weakest defense, especially when you're up against AI that can easily build up double your military strength in no time. Archers are a force multiplier- they are excellent when garrisoned in a city, and when strategically placed behind your melee units to make the battlefield their killzone, they make all the difference once the stuff hits the fan. If you expect to defend fast expansionism from the start, I'd go with 2-to-1 archers over warriors. They are just that good, when used right.

I will like to tell ya'll a story about three civilizations - Polynesia, Rome and the United States of America



From first glance, we see that America possesses the Statue of Zeus (a warmonger favorite), and that he is at war with both Polynesia and Rome (in addition to denouncing Rome). Thus, one might initially suspect that America was the warmonger who DOWed both of them. Crazy, isn't it?

A look at the endgame notifications paints a different picture:



T58 - Washington attacks Wittenberg, possibly in an attempt to enlarge his empire through conquest.

T62 - Kamehameha and Augustus Caesar declares war on Washington on the same turn, clearly a coalition effort

The exact details might remain a mystery, but we can definitely deduce a couple of things from this:
1) Wittenberg was likely under the protection of either Polynesia or Rome
2) Washington likely completed a wonder that was coveted by Rome
3) Augustus was just being an opportunist - the likeliest explanation was that he simply played off both Kamehameha and Washington, and when Polynesia was too entangled with America, his Legions struck (he did eventually DOW Polynesia, forcing them to complete the GW at T125)

Peng Qi said:

No? He's basically lost the game. His best bet is to try to make your situation as bad as possible in hopes people will declare war on you, then make a desperate final assault on you. He's literally a man with nothing left to lose.

Click to expand...

Don`t agree because the AI may not lose the game. There are only two other Civs on this world (India and China) too far to do anything in this medieval era and they don`t seem to care anyway. All Japan`s doing is killing itself.

A city in japan`s position should just be quiet and\or bribe another nation to go to war with me secretly so I don`t realise it did it (if it can afford it). that would be smart.

I decided to wipe out Japan for denouncing me straight after I let it live as it was stupid. I destroyed every Japanese city on the continent. Or so I thought. Luckily for Japan he had ONE city that I hadn`t discovered yet off map across the sea and I didn`t have the ship abilities yet to go abroad. So he`s out there somewhere, consolidating.

That`s one lucky nation!

I just hope they force a defeated AI not to denounce the victor who agreed to victory until a good few turns later in a patch. Certainly not next turn, it`s dumb.



It can still be logical and still surprise the Player though. Often a Player might think it`s not logical because he can`t see all that`s going on in the background. But for instance, during my war with Japan, I bribed China to go to war with it twice. Now to Japan it will look as if China went to war with it for no reason, but there was a reason, it just wasn`t obvious why to the attacked nation.

I think the AI is not too bad in general, most things that happen make sense eventually even if I don`t immediately know why.