Define AI

[Azi Tohak]

I know that many of us have our complaints about AI for RTW. However, I never really thought about what exactly AI is.

I am doing research right now on how to get a computer to tell me how far apart different items are on a circle, and then score those items placement versus a master. And it is @#*$@#( hard! I can’t program computers to save my life so I have to find what others have done, and I stumbled across the most interesting thread on what AI actually is. One of the people brought up the super chess playing computers, and how he thought they were AI (heck, so did I!). But another person told him that no, they were not ‘true’ AI capable of developing new thought processes, but they are merely huge, complex search programs. I thought that was interesting. It made me think about the total war games.

I played STW to death (my computer could not handle MTW until this new one). And while the computer was never really horrible (no ramming generals onto my spears, unsupported….), it also never really surprised me (that I can remember). I know it was not as easy as RTW, but I put that down to fewer units meaning less thought having to be done to figure out what to do with A as opposed to B, C and D. Now I wonder if maybe the computer was just better at searching, less programming having to go in.

Now for RTW…maybe the searching procedures were just neglected in favor of making it pretty. But that does make sense. I know how hard it is for me to get a computer to search for a bunch of pixels, I can’t imagine how bad it would be to program a foe on a 3-D battlefield with terrain, different unit types (a huge number really!).

That made me wonder, are there any games with RTWs tactical complexity (I think the strategic ineptness is inexcusable, I don’t see how the strategy could be harder than tactical!) but still have a solid search/AI or whatever you want to call it? I know everyone says MTW has a better AI, but I have not yet been impressed. It is not a moron, but it is not ‘good’ yet. Anything else out there?

Azi

[Productivity]

Not strategy, but I was surprised nonetheless.

I recently decided to buy Unreal Tournament 2004, and am playing through the single player game. Anyway, it's a capture the flag game and it is five a side, I have assigned one bot to offence, one to support (which seems to act like a midfielder) and two to defence. I myself roam.

So I grab the flag and it auto calls that I have the flag and where I am, and am running out of the enemies base with four behind me (and just about dead myself), when from around a corner my offence bot turns up and starts opening up with a chaingun. I though thanks, but put it down to pure chance. An AI offering realistic support. Never.

Anyway, after a valiant attempt to buy me time, he falls, and I'm again being chased by the opposition, this time at the bridge in the middle the support (midfield) bot does the same thing. At this point I'm thinking, well maybe this AI does work.

After playing for a bit longer, I've come to the conclusion that the AI does work. It does things that are smart. It attacks from multiple directions at once. If you call for support it will support you within it's sphere. The first defender is different to the second defender is different to the third efender etc.

Surprising and very nice to see. Also it seems Epic Games are the benchmark for supporting games, releasing mini-expansions for free to the community.

[Red Harvest]

I wouldn't worry too much about theorists that get caught up in the pedantic discussion of AI saying that chess computers are not "true AI." Give a chess computer a position it has never seen before and it will "score" it, make predictions of most likely moves, score the resulting positions, and select the move it has selected as best. It is good enough for games, and good enough to become World Champion. There are not many humans that even a good 20 year old chess computer can't beat. Yes, it fails to learn...although it is possible as the user to update the opening book to avoid the position in the future--or to call for a different move. The move choice can also be slightly randomized (this is done with the "opening books" already.)

Frankly, if any vid game could ever do half as good as my 20 year old Par Excellence chess computer (officially tournament rated as "expert" when it was sold) then I would be quite satisfied with that game's AI.

The real problem with today's games is in writing an AI (or rule set, whatever) that recognizes basic patterns, and responds appropriately--and sometimes with some random branching...that last little subtletly is often missed, but is the key to limiting human ability to ruthlessly exploit 100% repeatability of errors. RTW's AI responds 100% inappropriately in far too many situations compared to MTW. Quieteus likes to claim that MTW and RTW's AI are the same. I can point to things that show they are certainly not identical although they share characteristics. However, Quietus misses the forest for the trees, even if he turned out to be correct. His thesis only shoots a giant hole in his defense of RTW. RTW's battle dynamics and combat calcs, projectiles, and unit balance differ tremendously, so using the same AI would be the equivalent of fighting WWII with WWI tactics...or WWI with American Civil War tactics.

[Papewaio]

AI = artificial intelligence.

Games are more a series of look up tables. So it is like a menu of options.

If AI was high cusine, game bots would be McDonalds.

[Fragony]

AI = artificial intelligence.

Games are more a series of look up tables. So it is like a menu of options.

If AI was high cusine, game bots would be McDonalds.

UT AI is pretty complex, they actually work as a team. That can't be easy to program.

[doc_bean]

AI is artificial intelligence, this doesn't just mean it isn't intelligence because it doesn't come from a human being, intelligence is abstract. It means that it tries the best it can to be just like the real thing.

Unfortunately, we don't even know what the real thing is. So it's hard to imitate.

From what i understand (admitted, not very much) most modern games use a rule-base AI, essentially a series of if-then rules that can't change.

Now it is possible to go beyond this point, adaptable AI is possible in theory, but I think it is still too complex to use in games (processor demands).

Now about UT, I never played it, do the bots really work together ? Getting them to attack from different positions doesn't seem to hard (just assign a penalty if they get to close to eachother).

From my limited knowledge of 'intelligent' search algorithmes, it really does take an expert to use them correctly.

[Azi Tohak]

Alright Red Harvest, I understand that Mr. Chess Computer program is superior, but why is that? Why is it that now, CA (or many others to be fair) cannot include that little sublety for complex games like this? Would it be so hard to tell Mr. RTW AI
'if human has long pointy things in middle of main line
then send cavalry around to flanks'?

And by the way....didn't military commanders try to fight WWI with Civil War tactics? Or WWII with WWI tactics? Just saying...

Azi

[Productivity]

Now about UT, I never played it, do the bots really work together ? Getting them to attack from different positions doesn't seem to hard (just assign a penalty if they get to close to eachother).

I don't know, but I found myself getting pressured like I can be in a FPS by a good clan working together. They seemed to work well. As to if they are working together or using penalties, I don't care to be honest, getting simultaneously attacked from four directions is difficult to deal with.

[doc_bean]

Alright Red Harvest, I understand that Mr. Chess Computer program is superior, but why is that? Why is it that now, CA (or many others to be fair) cannot include that little sublety for complex games like this? Would it be so hard to tell Mr. RTW AI
'if human has long pointy things in middle of main line
then send cavalry around to flanks'?

How fast can the human turn its men ? Is there any backup he has that I (the AI)have to worry about ? Do I need to charge up hill ? Do I need to cav to guard my own men ?

There are hundreds of variables that have to be taken into account, humans can make quick estimations based on experience , and we can adapt.
Besides, AI takes up processor time, the game has to run smoothly.

And by the way....didn't military commanders try to fight WWI with Civil War tactics? Or WWII with WWI tactics? Just saying...

Azi

Yes, they lost horribly.
The French seem to have (had) a neck for using outdated strategy and tactics, which earned them their bad name when it comes to war.

[The Mad Scandinavian]

I concur about the UT2004 AI. All the UT games and a lot of FPS seem to have wickedly good AI's. And it doesn't seem as if it's because of bonuses. They just seem to know where to be to kill you and on UT2004 they do work together. Maybe it's just easier to make a competent AI for a FPS? Don't hold me to that. But this competency is really in contrast to a lot of strategy games. RTW is not alone in having AI problems.

Likes been said before, AI really is just a look up tool for if this then this. It can't be anything else since people can't program programs to write new code. Randomness just makes it seem real and can set the AI on a different path entirely, or not. There's a variety of other things all having to do with randomness such as battles lost that should have been won and Yucca Forests burned down. That's why little nations can sometimes rise to conquer the world. The more of these random events the program and processor has to deal with the more it has to recompute what is the best thing to do. Added to that is the amount of computations the computer has to make in the first place, such as move this unit stack from the city across this territory to the boat then send the boat to such and such a place, land, move across more land to attack an enemy army with the ultimate goal of attacking the city and taking the province. This compared to MTW where all that really had to be done was move this stack to this territory to fight the battle and siege the province. I'm sure it gets very complicated making an AI of RTW's caliber work. I sympathize with CA and if I worked for them I'd be on anti-depressants. I'd have probable started sometime around July or August last year.

[Red Harvest]

Alright Red Harvest, I understand that Mr. Chess Computer program is superior, but why is that? Why is it that now, CA (or many others to be fair) cannot include that little sublety for complex games like this? Would it be so hard to tell Mr. RTW AI
'if human has long pointy things in middle of main line
then send cavalry around to flanks'?

Two reasons it is superior: 1. Incredible amounts of development time by a huge number of brilliant folks--and extensive feedback and participation in the process by Grandmasters. 2. Frozen rule sets. The game rules don't change every year.

So how could TW series game AI ever compete with a human since it obviously can't achieve either of those? One way is speed of selecting the appropriate response. Unlike chess, the AI can order many moves without waiting for the human to respond (or in response to human moves.) On the battlefield the AI is not handicapped by the controls interface or the camera. For the human an interface is necessary, both for control and for the camera, some of this adds immersion. It sees things instantly and can respond nearly instantly. TW series battles have always been tough when action occurred in the trees...or spread out across the field...or when there was fog or snow that obsured vision (prior to RTW.) In RTW the AI has its best chances when there is a mad furball rush with armies close together.

Another way is to not make obviously stupid mistakes (something humans will do through oversight or misunderstanding.) Use a solid rule set, so that the AI can at least perform decently in a stand up fight. Sure, humans will find ways to game it, but it should take some effort to figure out how. Add to the rule set when easily remedied weakenesses are found (such as letting the human flank to take away high ground in RTW.) In RTW one gets the impression that the AI doesn't even understand how its pieces "move" on the battlefield. MTW had this to a lesser degree, but it was primarily an issue with its foot javelin units--and the horse javelins were deadly! RTW has the problem with most units. MTW seemed to understand basic use of its units with the exception of foot skirmishers.

Getting a solid rule set takes time. Players have to explore the combat engine, understand it, and learn how to use it proficiently so that the winning rule set can be developed. Then there will be inevitable tweaks to the initial combat engine to fix bugs and unintended effects...and another learning curve. At some point the actual game rules and unit type balance, must be frozen (not the same as the AI rules.) You can't have a good AI if the relative strengths of unit types changes fundamentally after the AI has already been completed. (Imagine if the Queen was "demoted" to moving like the King in chess--it would wreck all of the positional analysis scoring any time a queen was on the board. Same as if the knight was no longer allowed to "jump" pieces, or if pawns couldn't capture "en passant.") All the play testing and good human player "rules" must be fed back into the AI development. How do you modify the AI to emulate what a human does so well? That is the difficulty. It takes time, lots of input, and some talented programmers to develope a strong AI.

I suspect CA has some quite talented programmers for AI, but I've not seen indication of the committment from above to get RTW up to snuff. AI development/tuning/tweaking must be extremely schedule dependent and much of that load is going to be on the backend once the game itself is nearly complete--the combat engine must be near finalized anyway. I have not seen evidence of the necessary level of playtesting with RTW, nor has the schedule been one that would allow good AI to be developed. Unit balance is fouled up so that counters don't work that well, the charge is mainly based on armour level, the combat engine encourages furballs, and the patching was cut way short of what was needed.

And by the way....didn't military commanders try to fight WWI with Civil War tactics? Or WWII with WWI tactics? Just saying...

Yes, at the start of each war many/most are using the tactics of the previous war or at least doctrines based on what are perceived to be the lessons from the previous conflict. The commanders either adapt quickly to the new reality, or get trounced and are remembered as history's goats while new commanders take their place.

[The Mad Scandinavian]

It would be really cool to see a programers viewpoint on this. Nothing too complicated but just the theories behind what they do to make the AI work. If that's not possible to explain simply then I'll shut up.

[Nelson]

Chess enjoys a very real advantage for AI developers as opposed to games like Total War. That is not to say that AI for chess is simple. It surely is not. But chess is far more discreet with regard to potential moves on each turn. Every game starts exactly the same way. The OB for each side is identical. Each piece can only get to a few spaces on each turn if it can move at all. Chess AI can therefor be expected to weigh anticipated counter moves.

In Rome, each battle is different. The units can be radically varied. All units can move at a walk OR a run IF they move at all. Units can change formation. Some can shoot or melee. There is terrain to consider. Knights don’t get tired in chess like they do in Medieval. Rome AI must determine what to do with an army consisting of from 1 to 20 units of who knows what continuously in real time even as circumstances change with regard to morale, fatigue and terrain. Introduce tired knights, frightened pawns and undulating chessboards with forests and rivers and then we can compare the two games more evenhandedly regarding AI.

[Mikeus Caesar]

Well, when you think about it, present day game AI isn't really Intelligence. It's just a bunch of programs and code that has bee designed to mimic intelligence. Just look at the creature from the game Black and White for example. It appears to be like AI, but if it was AI, it would learn. It would learn how to do thing for itself, rather than having you teach it, or should i say, program it, not to do or to do things by petting it or hitting it. It has limits, and true AI wouldn't have limits. The limits for it's expansion and learning would be, quite frankly, limitless. True AI would learn from it's experiences without human help. And, going into the realms of Sci-fi, in the future, true AI will be self aware.

Define AI

Anything that isn't from RTW....

[Productivity]

What I think would help the Rome AI (at least defensively) is if it just picked a tactic, and stuck to it. I am playing as Spain now, and the Scipii reguarly siege Carthago. Now every time I sally out, they withdraw to a point away from the walls, and form up into formation. But on the way there, they may try to re organise 3-4 times, often just as I am about to hit them.

If the AI could come up with a decent defensive formation, and then hold it, it would do far better, rather than me hitting it as it is re-arranging. Maybe give units a factor which determines how "key" they are in any formation, so cavalry would have a low value for this and be able to mvoe around at will, skirmishers would have a somewhat higher, but still low value, and key heavy infantry would have a very high value.

[Papewaio]

UT AI is pretty complex, they actually work as a team. That can't be easy to program.

Its still just a look up table not real intelligence by any means.

= Main Order Program Menu=
Take order,
smile at customer,
enter order,
smile at customer,
pick up order,
smile at customer,
give order,
smile at customer.

- Take Order Sub Menu -
What would you like today?
.
.
.
Supersize that?

- Make Burger Sub Menu-
.
.
.
Put Burger on Bun,
Smile at minimum wages,
Put Pickles on Bun, /* Note if no pickles on bun it is no longer a main meal and becomes a confectionary item */
Smile at minimum wages,
Squirt on Cream [Check CreamVar is Mayo > CreamOfSomeYoungGuy],
Smile at minimum wages,
.
.
.

====

Same for the UT bots they are given a destination and target type range. They don't learn, they use SQL like tables. So at best they should be called primitive AI where AI is artifical instinct.

[Productivity]

Same for the UT bots they are given a destination and target type range. They don't learn, they use SQL like tables. So at best they should be called primitive AI where AI is artifical instinct.

While that is true, I don't think anyone actually cares, if it is providing the appearance of AI. They have been programmed to have a diverse enough array of instinctual responses, that when put into a situation and adding randomisation, they appear to be intelligent.

True AI would be vastly difficult, and I think the closest anyone has come has been adaptive adjustment (a racing game Powerslide used this, it would map it's own routes, acceleration etc. across the course, and map yours. It would take the stretches of course where you beat it, and move it's new course more towards yours). However that's still not AI, it's still working off defined equations. When someone can work out a program that allows it to rewrite it's own equations, then you will have AI.

Rome probably could use adaptive patterns, if it tracked the battle, and worked out when it was winning the battle, and when it was losing, and then building on those strengths and eliminating the weaknesses. So if I was playing spain, and I allways took a hammering from the AI's archers, over a campaign it would slowly start to build more and more archer heavy armies, until it started getting better returns off of other units. Essentially using marginal returns to units.

[doc_bean]

I think a few people here are overestimating actual intelligence.

We mimic what we see other people do. We 'build' certain patterns in our brain. Our thoughts don't appear out of no where, they are formed by a system trained by years of conditioning.

[Azi Tohak]

Seriously, I would love it if the computer would just pick something and stick to it. For both defense, but especially defense. I would love to have one massive wall of infantry slam into each other, but the way the computer works it is a whole bunch of little pieces. I pick on a piece, crush it, rinse and repeat. If the computer would use its phalanxes just as one wall, it would be far harder!

But on another note, I think it would hilarious to see RTW try to make a sandwich...Sorry...that was just a funny mental picture I had...I will go back to work now...

Azi

[Papewaio]

I think a few people here are overestimating actual intelligence.

We mimic what we see other people do. We 'build' certain patterns in our brain. Our thoughts don't appear out of no where, they are formed by a system trained by years of conditioning.

True we do mimic, but that is in itself a sign of intelligence. How hard would it to be if RTW had Hard - VH then Mimic level.

In Mimic level it mimics your style of play and picks from the best styles given its current formation. It also learns from campaign to campaign what is the best tactic given a situation.

Even if the Strategic Mimic was simply an order of buildings and then troop mix it would be a step up.

[Azi Tohak]

Mmmm...mimic level...

See, I'm just so darn good, it would be the unstoppable force meeting the immovable object

Something like that. But anyway, seriously, wouldn't that be so hard to program? Given the variety of terrain, I have to change my strategy. To be sure, it does not change by much, but it seems to me that would be awesome. And terrible at the same time. I might actually lose a battle to the AI!

Azi

[Papewaio]

Mimic the order of build style wouldn't be too hard, even if it just decides to proportion the same amount to barracks/ports/market as the player.

Battle style may be more difficult as then you would need probably some sort of neural network, then again maybe just having a wider range of tactics and having a history of which work best against the units you have deployed might be good enough.

Charge general across bridge won 1 / Lost 99
Charge Naginata across bridge won 3 / Lost 16
etc
The computer orders figures out which tactics worked the best in the past and randomly selects from the current best 3 or so. That way the player is up against a hard opponent but doesn't know what they are going to do.