Best PC Game AI Ever - Maybe Firaxis Can Hire Their Coders as Consultants?

[Aristos]

It's a bit optimistic to see that sort of evolving AI in a CIV game any time soon.

specially with the current development team...

Soren, on the other hand... maybe, just maybe...

 

[Mark the Bold]

So Jharii, if the code was right, and some super-server had say 10 random though persistantly coded memetic AI's playing against random human Civ V players continuously input into the server. Over long enough time one could develop some almost perfect AI that could counter anything thrown at it? Or should I say, it could counter any "human" behaviour thrown at it? The code would almost write itself. The difficulty I feel would be in the code itself (obviously) in determining what human action relates to its goals and how the AI would manage to 'delete' old inefficient behaviours. But I see the logic. Fascinating stuff. In fact, future of human technology stuff. I need a cigarette....

The really funny thing is that THIS is what I would pay DLC prices for. Heck, you can keep the downloadable civ. Give me the "Deep Blue" AI code upgrade. I'd gladly pay extra for that than another Civ.

 

[AriochIV]

SSG's Warlords series had very good AI for the time.

 

[Mark the Bold]

For most games the level of the AI depends more on how easy it is to make an AI to it than on how well they did it. As mentioned earlier it's not hard to make an AI which plays FPS at a very high level because you can simply increase their accuracy. Games that are formulated as an optimization problem can often be optimized more efficiently by a computer than it can by us, and combinatorial problems can be crunched more efficiently, but with a more complex game like Civ 5 which requires a lot of judgement it becomes harder.

Alright, alright I wave the white flag on the Perfect Dark comment. Just trying to throw an idea out there. Maybe my cartridge had some Deus Ex Machina upgrade, but what made the Perfect Sims great in my book was that teams of four of them would split up and surround my teams position. That's all I meant in great AI. Still better than so many shooters put out today IMO.

 

[Mark the Bold]

SSG's Warlords series had very good AI for the time.

Yes. However, I would find that the AI would get bottled up in choke points and have like a hundred units stuck unable to attack. Well at least on Warlords III (the only one I played)

 

[Jharii]

So Jharii, if the code was right, and some super-server had say 10 random though persistantly coded memetic AI's playing against random human Civ V players continuously input into the server. Over long enough time one could develop some almost perfect AI that could counter anything thrown at it? Or should I say, it could counter any "human" behaviour thrown at it? The code would almost write itself. The difficulty I feel would be in the code itself (obviously) in determining what human action relates to its goals and how the AI would manage to 'delete' old inefficient behaviours. But I see the logic. Fascinating stuff. In fact, future of human technology stuff. I need a cigarette....

The really funny thing is that THIS is what I would pay DLC prices for. Heck, you can keep the downloadable civ. Give me the "Deep Blue" AI code upgrade. I'd gladly pay extra for that than another Civ.

LOL. Yes, I've drooled over this stuff for several years. And yes, it would write itself. Not only that, but the difficulty levels would write themselves.

I can build Stonehenge in 75 turns.
I can build Stonehenge in 60 turns.
I can build Stonehenge in 50 turns.
I can build Stonehenge in 40 turns.

You are correct regarding the difficulty combining the logic with the game mechanics. The functions are all there. We use them all the time, as does the AI. The goal would be change how the AI views its strategic and global resources and victory conditions (needs) vs. everything else in the game (wants) and how to apply the functions of the game to achieve these.

What is really interesting is that you could give an AI city this same type of AI so it managed itself more efficiently over time.

The game itself is complex enough. But if they designed it, they could actually have the game play itself X amount of times to establish the base AI and then unleash it to the masses. Otherwise, we would be playing games in the beginning with virtually no actual AI. Think of 1 million monkeys with 1 million typewriters, living an eternity. Eventually, one will reproduce the entire works of Shakespeare to the letter.

Sure, you could kickstart it with some basic instructions and meme chains. But man, it would be so much more interesting to see it just build itself from the ground up.

 

[UknowsI]

You are correct regarding the difficulty combining the logic with the game mechanics. The functions are all there. We use them all the time, as does the AI. The goal would be change how the AI views its strategic and global resources and victory conditions (needs) vs. everything else in the game (wants) and how to apply the functions of the game to achieve these.

Do you mean this should only be used for the macro AI or also for detailed management? I think it could work with on a macro level, but the biggest flaws in the game are on a more tactical level. Doesn't matter if the AI triggers "Attack Jharii with 73% of your land units at his left flank" at the perfect timing if all he does is to suicide his units.

 

[Mark the Bold]

Do you mean this should only be used for the macro AI or also for detailed management? I think it could work with on a macro level, but the biggest flaws in the game are on a more tactical level. Doesn't matter if the AI triggers "Attack Jharii with 73% of your land units at his left flank" at the perfect timing if all he does is to suicide his units.

I agree UknowsI. The beauty of 1upt is that there is a finite amount of unit location situations. Sure it is a big number, but easily manageable by a computer. For instance I have 1 horesman, 1 spearman and 1 archer and comp has the same in a say, 10 x 10 tile area. There are only so many attack patterns and defensive situations available to either. Computer should be able to determine "if horseman does x than I should do y" on my turn, etc. etc. And the Deity / Emperor characteristics of difficulty should be based on how much CPU power is allocated to these thoughts / reactions. Yes, yes I know the terrain makes a difference, but really all the computer has to do is figure out attack strength value vs defense strength value in a spreadsheet so to speak. Sure we see pretty tiles and mountains, but the comp should only see numbers. It should easily be able to be more tactical than currently. The code is already there I'm sure, just needs a little TLC from some coders.

 

[Soro]

I found the EU3 AI to be horrible, maybe worse than Civ 5. There is a while since I played it, but I remember at first I really liked the concept of the game, but that I got bored because the AI was too predictable. After a while you learn the how the AI works and it feels like a set of triggers mechanisms with a small RNG attached to it.

When was the last time you played Europa Universalis III? The game has been repeatedly patched. At the start it was truly bad, but this far in it's very, very good. And not just on combat, either. It can form interesting alliances, and even negotiate reasonably well at the conclusion of wars.

Personally I think the tactical combat AI in Civ needs help from chess programmers. The tactical AI in Civ would be greatly increased if a unit could do what was smartest thinking no steps a head and only considering tiles 3 steps away. If they made it able to think 1 or 2 turns into the future and 4 tiles in each direction the game would be changer completely.

Would that were possible! -But chess AIs are so good because of enormous databases in the ability to look at them while the other player is making his/her move. Civ (and other strategy games) lack the databases, as well as the ability to look at analyze when it's the player's turn. Sadly.

 

[Jharii]

Do you mean this should only be used for the macro AI or also for detailed management? I think it could work with on a macro level, but the biggest flaws in the game are on a more tactical level. Doesn't matter if the AI triggers "Attack Jharii with 73% of your land units at his left flank" at the perfect timing if all he does is to suicide his units.

Hmmmmmm. Both?

Game management would definitely be a lot easier. But I didn't initially think about the tactical side. My guess would be that each game AI would have two separate AI's, one for game management and one for tactical management. This way it could separate the priorities, allowing the tactical AI to create wants and needs for the game management AI to fulfill (and vice versa, where necessary). Once the units are created by the game management, they are passed back to the tactical management.

It may even be necessary to provide this same AI to each unit to interact with the tactical management. I don't know, immediately. It's a lot of theorycrafting as to how it applies to Civ, and like I mentioned, some of the concepts have been just out of my reach.

 

[Sox]

When was the last time you played Europa Universalis III? The game has been repeatedly patched. At the start it was truly bad, but this far in it's very, very good. And not just on combat, either. It can form interesting alliances, and even negotiate reasonably well at the conclusion of wars.

As much as I love EU3, i don´t think i would call the AI good... Which leads me to the OP.

There simply isn´t any good AI in games, there are only mediocre and bad.

As much as i love the optimistic tone of some of the posters I personally don´t think the AI in Civ5 will improve that much. I just think that the tactical (1upt) and strategy (Civ) parts is, today, to hard to combine. Prove me wrong 2k!

It does however bring a smile to my lips when I see posts about AI (not game) and people saying that "hey we can all come together and create the greatest AI ever through mods!" The real research about AI is leaps and bounds above the AI we are seeing in games today and while modders surely can, and do, improve upon the AI in many (most?) games I fear Civ5 is a lost cause.

 

[Olleus]

I'd suggest two other titles for consideration:

1) Distant Worlds.

2) Europa Univeralis III.

Both have excellent AI.

Never played Distant Worlds, but surely you're joking about EU3? The military AI is leagues worst than the Civ5 one and the diplomatic AI just seems to hate you pretty much no matter what.

 

[Jolly Rogerer]

....

Would that were possible! -But chess AIs are so good because of enormous databases in the ability to look at them while the other player is making his/her move. Civ (and other strategy games) lack the databases, as well as the ability to look at analyze when it's the player's turn. Sadly.

True. And even with a database and the willingness to use it during the players turn the massive (and unique) map, the fact that all units can move every turn, the element of chance in combat, etc. really limit the capacity to look deeply at the situation. I wonder how many possible moves there are in turn 100 of the average Civ4 game?

 

[UknowsI]

When was the last time you played Europa Universalis III? The game has been repeatedly patched. At the start it was truly bad, but this far in it's very, very good. And not just on combat, either. It can form interesting alliances, and even negotiate reasonably well at the conclusion of wars.

As I mentioned, it is a long time since I played it. If the AI has improved, then that's great, but it was horrible when I played it.

Would that were possible! -But chess AIs are so good because of enormous databases in the ability to look at them while the other player is making his/her move. Civ (and other strategy games) lack the databases, as well as the ability to look at analyze when it's the player's turn. Sadly.

I'm not talking about a DeepBlue style AI. Chess AIs from the 80ties are better at combat than the Civ 5, and if they are looking 1 turn a head they wouldn't need to look at all possible outcomes but have a simple evaluation function for the state. But as I mentioned, even looking 0 turns into the future would be an improvement if they were able to evaluate moves and positions better. I have seen situations where the enemy has their artillery in a good position to shoot and kill my unit while they instead chose to move into the grassland tile in front of my rifleman... I can't even imagine how it's possible to program an AI to consider that a good move.

True. And even with a database and the willingness to use it during the players turn the massive (and unique) map, the fact that all units can move every turn, the element of chance in combat, etc. really limit the capacity to look deeply at the situation. I wonder how many possible moves there are in turn 100 of the average Civ4 game?

It is of course impossible to evaluate all possible moves, but it should be possible to evaluate the most realistic end states in a local area. Giving bonuses for damaging enemies, giving penalties for exposed ranged units and so on.

 

[Jharii]

I can't even imagine how it's possible to program an AI to consider that a good move.

That's exactly why I think that each unit would have to have it's own AI. Then they could be catered to the unit type. If they were memetic, they would even "interact" with each other.

 

[TheMeInTeam]

Warcraft III AI was very very easy to defeat, especially when fighting orc or human AIs which would quit on harvesting resources to attempt defense even against basic harass . It had excellent APM but wouldn't make great actual micro decisions. I could beat it easily and soundly and was nowhere near the level I am at in SC II for example. SC II's is much better, but no where near unbeatable.

 

[Benzidrine]

specially with the current development team...

Soren, on the other hand... maybe, just maybe...

Soren programmed the vanilla AI of civ 3 and civ 4, I am not sure he is the expert we are looking for.

 

[Psyringe]

I think the key point has been made in one of the early posts: To have a good AI in a TBS game, you need (at least to some extent) to build the game around it. GalCiv2 did this and succeeded. It also made the game less enjoyable for some players because its game design consistently only included features that the AI could handle reasonably well. Older games (SMAC or MoO2 for example) tend to give the player more strategical options and variety, but their AI ignores half of them.

That said, I think that Civ4's AI is actually a higher achievement than that of GalCiv2, because the AI for Civ4 was much more difficult to write. Taking this into account, Civ4's AI - even in vanilla - is an impressive feat, although the hardcore players have of course analyzed it to a degree that all its weaknesses are clearly visible (and quite annoying) for them.

However, even in Civ4, the game was partly designed around the AI - though not as radical as in GalCiv2.

 

[uberdubersoldat]

chessmaster

 

[Psyringe]

chessmaster

Not sure if that was meant as a joke, but the approach of chess AIs (scripted beginning, brute-force midgame calculations with look-ahead, and a combination of both for the endgame) would utterly fail in Civ. While the AI would come up with a good move (in unmodded games), you'd have to wait a couple of days or weeks until it completes the necessary calculations.