I didn't buy any of their expansions for RTW after visiting the forums. Mainly becuase the AI is still stupid and I know i will get sick of the game after 2 to 3 days.Quote:
Originally Posted by Satyr
Printable View
I didn't buy any of their expansions for RTW after visiting the forums. Mainly becuase the AI is still stupid and I know i will get sick of the game after 2 to 3 days.Quote:
Originally Posted by Satyr
I suppose CA could create a really competent or 'killer' AI, but there are - in my opinion - a few simple reasons why they don't:
1 Users;
2 Profit;
3 Time.
And in that order.
Most users will be either relatively new to the game or won't have a real powerfull system on which they run the game. I mean the more AI you put in the game the more it will depend on users with a powerfull system since the game will ask more from it. Not everyone is going to buy a new computer just for one game... Henceforth CA will have to either limit the number of aspects which effectively will limit the number of troops the AI can use or CA will leave the AI relatively straightforward as it is. Clearly there will be improved AI in the new TW versions, but it won't be 'killer' AI - it won't be anywhere near human way of handling tactics.
Profit is simpler to explain: what does attract new customers? Yep, graphics. AI is a nice bonus, and will secure the more experienced lot. But still a fair amount of the veterans will still buy the new game - since they too love the graphics part. (Check the RES GESTAE mod subforum in case you don't believe me...)
Time also limits the amount of work that can be done on AI. CA has to make a certain amount of profit within a certain time limit. They are an commercial enterprise, we all know this very well. This also means you've got to balance expenditure and income. Henceforth you can't just take all the time you want to create your game as you will need income to do so, income which mostly depends on the sales of the new game. Only new releases are really in the big time - this is why most games are shortly followed up by an expansion. This has little to do with 'now we have extra time, the first was simply a sort of demo version thingy' but rather more with 'gamers are going to want more of the first, and new gamers will buy the first because of this expansion pack thingy'. This is also why RTW needed such a large amount of patches, and why the were such ridiculous flaws in the AI such as missile unit's charging melee ones.
Hi everybody,
I've been around a while but mostly as a lurker. This is something I've spent some time on, and I have perhaps a unique (if not altogether insightful) perspective on it.
I am the lead designer of FreeOrion (www.freeorion.org). One of our models for 'successful' games we'd like to, uh, imitate (read: borrow/steal) things from is Total War. Obviously, we're in space, so it's not a clean transition, but a lot of the issues are the same. Strategic engine combined with tactical engine. We are currently in the process of designing the tactical engine, and starting work on the AI for both that and the strategic engine.
From the very beginning of our project, we had people clamoring for AI brilliance. They said what a lot of you are saying. We don't care about graphics (as much), just give us a challenging game! And a lot of us were in that camp as well.
The simple fact of the matter is that AI is the single most difficult aspect of game design, and the biggest time and money sink in any sufficiently large project. Our expectations of game AI are pretty high, and for good reason. Our games have come so tremendously far over the last fifteen years in almost every category, we want AI to have made similar progress. We even say 'stop spending so much on graphics, just give us AI' because of the huge discrepancy.
The fact of the matter is that, when we develop an AI capable of beating really, really good humans at Chess, it makes world news. And Chess is very structured, such that almost every operation can be reduced to math, proabability, and branch prediction. To say that Total War or any other modern strategy game isn't even close to that level of structure would be a huge understatement. We just don't have the level of understanding in this industry yet to reliably produce intelligent AI that is really satisfactory. The route most developers have taken amounts to an enormous list of weighted scripts that can be reduced to 'if this happens, do one of the following things, and if nothing happens, fall back to this behavior'. To my knowledge, only the GalCiv developers have taken the alternative route and tried to reduce all of their AI operations to something a CPU can just sit there and churn away at, such that the more CPU you throw at it, the better it is. That's something they've been after for a long time, and it's admirable - but it's also very difficult, and they have yet to demonstrate it on any kind of realtime basis, since the whole idea behind GalCiv's AI is that it ponders while you ponder. Battle AI can't do that.
All of that aside, I'm not trying to apologize for CA. None of this matters to the consumer. You and I want an awesome AI and we want it now, and we can recognize when it's broken. Doesn't matter why. But it is important to understand the problem that we (and they) are wrestling with. Fishpants already informed us of what should have been painfully obvious: the graphics folks aren't the same people as the AI programmers and it is never an either/or consideration. Like any company, CA has to follow certain business practices to remain afloat: they've got to generate excitement about their product, generate the product, and sell the product. The most efficient way to generate excitement is to release cool-looking screenshots and do interviews. Demos are always a dicey prospect, since every minute they spend putting together a demo is one less minute spent working on the final product -- and that IS an either/or proposition. Announce a demo and you've just lost dozens, if not hundreds of collective hours of development on your product. So they can't just 'show off the AI' at the drop of a hat, even if they wanted to.
Frankly, the only way AI will ever improve is if we continue to support companies for whom it is in their best interest to improve it. That means buying GalCiv and buying M2TW and trying to figure out how to make them better - something these forums do quite well. But it's foolish to suggest that CA doesn't care about AI, or that it's not a priority for them. It's just a fact that AI is behind every other aspect of gaming today, industry-wide, though there are several noteworthy exceptions that make the future look pretty exciting for everyone involved. I wouldn't be surprised if, one day, we end up buying AI-processing add-on cards just like we have GPUs today. Because truly competitive AI will require a fundamental shift in the way we think about it and in the way it's done in most mainstream products today.
Cheers,
Aq
You make some good points there, but I don't think most TW fans are asking for almost human AI. Most of us would be extremely pleased to see M2TW have a decent level of AI competence that was shown in Medieval and Shogun, which shouldn't be impossible.
That was great, Aquitaine. Thank you for coming forward.
Consider the RTW strategy map in that context.Quote:
The fact of the matter is that, when we develop an AI capable of beating really, really good humans at Chess, it makes world news. And Chess is very structured, such that almost every operation can be reduced to math, probability, and branch prediction. To say that Total War or any other modern strategy game isn't even close to that level of structure would be a huge understatement.
The Medieval: Total War map had provinces, which might be bordered by three or four other provinces. Each army had -- at most -- four or five options for movement, including the option of sitting still.
A stack in Rome Total war has literally hundreds of possible destination points to choose from.
To a layman, it appears the AI faces a vastly greater challenge on the RTW map than the MTW 1 map.
So, to give the whiners some credit, it appears you could do that with the strategy AI. It appears that each faction flashes through its options after the human player completes all his moves. If the AI allowed, say, twice as much time to sift through its options, it could be engineered to “ponder.”Quote:
We just don't have the level of understanding in this industry yet to reliably produce intelligent AI that is really satisfactory. The route most developers have taken amounts to an enormous list of weighted scripts that can be reduced to 'if this happens, do one of the following things, and if nothing happens, fall back to this behavior'. To my knowledge, only the GalCiv developers have taken the alternative route and tried to reduce all of their AI operations to something a CPU can just sit there and churn away at, such that the more CPU you throw at it, the better it is. That's something they've been after for a long time, and it's admirable - but it's also very difficult, and they have yet to demonstrate it on any kind of realtime basis, since the whole idea behind GalCiv's AI is that it ponders while you ponder. Battle AI can't do that.
But, as you point out, that would be a very expensive process.
Unit movement on the strategic map is not as complex as it seems though because of the various choke points and the fixed positions of the cities. Unit movement is much more complex in CIV and that has always been good but improves with every iteration of the game (it should be noted though that the calcs for AI unit movement are by far the most CPU intensive aspect of the game).
Thank you for that post, Acquitaine, I'm sure we all appreciated that. :2thumbsup:
Most of us do understand that CA is indeed a commercial company with bills to pay and shareholders to satisfy, and the fact that profit is and always will be the number one priority. We accept that, and we appreciate what CA has done as far as Total War is concerned. No other game can come close to the kind of experience Total War offers, and the direction that the company chose for the series is, IMHO, a great one. :2thumbsup:
While the majority of us aren't programmers, we at least understand (despite our limited knowledge) that making an AI that can think as good as a human is, at least for now, impossible, and the best that programmers can do is to tell the computer to "If this is A, then do B, or else do C", something to that extent. I daresay that what most of us are really asking for is not an AI that can do wonders on the battlefield... no, I'm not asking for that. It's too much to ask and we know it.
What we want is an AI that seems great at least at face value... one that doesn't frustrate you because it barely put up a fight... one that doesn't make you pull out your hair because it did something that you'd NEVER do in a million years (ie. charge your general alone into a mass of formed infantry; make your phalanx turn 180 degrees 2 meters in front of your troops, etc). At least give the AI certain absolute or near absolute DO'S and DONT'S so that it doesn't do stupid things... now I feel that this, at the very least, can be done. You might think that it might make the AI predictable and easier to beat... but I don't think so. Example, you could tell the AI to "only send missile troops to melee if: out of ammo, not the only unit engaging the target unit in melee"... something like that. That's what most humans would normally do, right? I'll be very happy with that because at least it looked as if the AI was actually thinking.
I don't want unbeatable AI (that's not fun :no:), I just want it to seem to be smarter, I hope that that can be done. :2thumbsup:
Again, thank you for your input. I do believe that the future of Total War and PC single player strategy gaming AI is bright, it'll all be a matter of time. Good luck to you and FreeOrion. :2thumbsup: :2thumbsup:
I'm not an AI programmer, but I don't think you can reduce RTW's strategic map complexity just because of choke points. The choke points were probably put there (among other reasons) to reduce the load on the AI, even if only a little bit, but that doesn't do a lot to alleviate the number of factors involved in how a human moves an army on the strategic map:
THE TOTAL WAR AI, by Aquitaine
- MTW: Rebels. Fight!
- RTW: Do I ignore those rebels, or move this small stack out of a city to deal with it? Do I have enough intelligence on the rebels to know if my small stack can handle them? Or do I wait around for a returning army to clean them up on the way back? Can I weigh appropriately the danger of exposing my city versus the limited intelligence I have on neighboring enemies?
- MTW: I have a navy, I guess. That means I can make this trip in two turns instead of seven. Hooray linear pathfinding!
- RTW: I have a navy. Wtf do I do with it?
- MTW: I want to attack the Uesugi. I border them in exactly two places, so I should look at the troop counts in both places and then attack. Maybe I'll weigh the difficulty of assaulting the terrain in each of those provinces, but maybe I won't.
- RTW: I want to attack the Gauls. Setting aside for a moment that my grandmother could attack the Gauls without any help, do I head directly for the nearest city, or do I spy a large Gaul stack and move to ambush them near this choke point? How many turns can I safely lay siege to their city without exposing my flanks?
- MTW Diplomacy: I hate everyone. Hey, let's have an 'alliance.' Ha, ha. No, really.
- RTW Diplomacy: I have a migraine trying to figure out how I feel about you. Give me ten thousand florins or ducats or something to make me feel better.
...I could go on. We aren't talking a little bit more complex. We're talking an order of magnitude. But again, this doesn't forgive obvious AI blunders. It's one thing to go 'well, it's hard! sob. poor CA.' It's another to go 'when the player can spot an obvious flaw, it ruins immersion, so we should complain about that...while keeping in mind that it's unlikely any other developer would have done better.'
I'm not really talking about a human-level AI. I'm just talking about an AI that can give a human a decent challenge. Not the Deep Blue of Total War AI.
I wouldn't even know where to begin on a battle AI. Even on a strategic map, the number of decisions and variables is pretty significant but can, ultimately, still be weighed. On a battle map, any decent TW player is considering and rejecting a huge number of decisions every second. To you and me, one glance at the screen tells us more about the situation than the computer can figure out just by comparing x,y, and z values of the general versus the values of the phalanx. In one sense, there is so much information available to the AI on the battle map that I don't have any idea how it prioritizes. It would be very easy for the AI to simply hold back the general nearly all of the time, just as it's easy for it to aggressively use the General. It's exrcruciatingly difficult for it to know exactly when to commit the General, whereas, for you and me, that moment is often the turning point of the battle -- the 'hold it, hoooooooold it .... now!' The AI probably knows, okay, the general is over here, and the general's unit feels this way about infantry and this way about spears and this way about cavalry, so we know we probably would have better luck charging this guy than this guy ... maybe we should do it now?
When we look at the screen, we can do almost instantly what the AI is largely unable to do at all - pattern recognition and intuitive planning. Being able to predict the exact point of weakness in your line and knowing not only where, but when to reinforce it, is just something we do a lot better than a computer can.
But look on the bright side. At least we won't be conquered by robots any time soon.
AI can begin to form patterns is the problem. Some of my friends found that when fighting me on LAN. They had 'perfect' strategies which always worked against AI, but when facing a human player you are never sure of how they will react.
Even when modding teams revise the AI there are still patterns and logic to the processing. It may not become apparent for a while but eventually you will find a sure way of defeating the AI. And there is also making the AI fun to fight, not ultimate.
There is a thin line between a push-over and an un-beatable enemy. It will be almost impossible to get it in the middle, because all players are different and play differently.
I don't know how anyone could make an unbeatable AI, that didn't cheat. As in the Hojo horde, or a massive morale bonus. I would love to see an AI that would try to flank me, or attack in echelon, or maybe just attack with a solid wall of infantry. I think it is pathetic how the AI uses Phalaxes in RTW. They should forever stay a wall of points. Wheel, advance, about-face... that is all they need to do. But they don't. One unit goes here, another crosses its path, and now you have a massive jumble of pikes that a swordsmen devour.
The strategic AI... now THAT would be the fun part. I understand how that one is appalling complicated. But... I suppose I wouldn't mind a little bit of cheating in terms of intel for the computer. But none of this in debt crap! Darned Hojos.
Azi
I believe that most people who had enjoyed playing against tactical ai from both STW and RTW were asking the ai to be equally smart or as cunning as a human ai. All they want is an ai, that knows how to maintain a proper and decent formation when approached by or engaging the player's army.
Simply by maintaining a proper formation when defending or attacking, the ai can present a pretty challenging tactical game. Something that you can find in the previous TW series but not in RTW.
Another thing that is more aparent in RTW is the bad combination of troops type you find in AI armies during campaign games. Considering that multiple AI armies are stack in a single province on a "Risk" style campaign map in previous titles, the only time the player face a bad AI combi army were during the early stages of the campaign. I wonder if CA would make the AI form a 'proper stakes armies' before sending them to battle.
My hope would be that M2TW AI;
1.)Maintain a proper formation when defending or reacting to players movement. (and not the famous archer charge and stray pikes chasing cav routines)
2.)Form proper armies before setting off to attack anyone. (not small sucidal expeditions or peasants & 'fodder units class' stacks) I think this aspect is very important with regards to the RTW type campaign maps.
3.)Make use of the maximum range of their missile units.
It is certainly true that it is difficult to make an AI that can compete. Especially since so many things are connected. Errors on the campaign map lead to armies that easy to defeat on the battle map. Even if the battle AI was very good, it would still lose because of the piecemeal way armies are often shoved over the strategic map. However, the AI could better do with a few simple rules. Of course no simple rule can apply to all situations, but some rules are that useful that you're better off following them blindly than not following them at all. An example would be: don't charge your general into a phalanx!
ROFL. Brilliant.Quote:
Originally Posted by Aquitaine
This is a very good point and exactly what troubles the CA programmers so much. How can you turn information into numbers, and how can you prioritize them? But there are a lot of simplifications that can and must be made, in order to make a decision. Perhaps CA is aiming too high here, leading to complex decisions which can (and do) go either way, which is always wrong in battle.Quote:
To you and me, one glance at the screen tells us more about the situation than the computer can figure out just by comparing x,y, and z values of the general versus the values of the phalanx. In one sense, there is so much information available to the AI on the battle map that I don't have any idea how it prioritizes.
But the AI can figure this out, by for example looking at morale fluctuations. If a units morale drops below 0 repeatedly for 3 combat rounds, it routs. If you as an AI register that a lot of morales of your opponents units are sometimes dropping below 0 (and back up again), it's time to commit that general. Until that time, set GeneralAttackOrder = 0.Quote:
It would be very easy for the AI to simply hold back the general nearly all of the time, just as it's easy for it to aggressively use the General. It's exrcruciatingly difficult for it to know exactly when to commit the General, whereas, for you and me, that moment is often the turning point of the battle -- the 'hold it, hoooooooold it .... now!'
It just takes the programmer time (game development, money) & resources (like, how often do you want the AI to check this condition? we only have so much CPU cycles).
Yeah this is considered. Too bad cavalry still manages to charge spears head on. Again: Simplification is in order, CA are perhaps trying to implement too much of Sun Tzu's Art of War, without getting basics right, first.Quote:
The AI probably knows, okay, the general is over here, and the general's unit feels this way about infantry and this way about spears and this way about cavalry, so we know we probably would have better luck charging this guy than this guy ... maybe we should do it now?
That is all we ask.
I think what you are asking is very reasonable. Even given a decent understanding of the complexity of the AI, it seems like CA bit off more than it could chew making the jump from MTW to RTW.
I guess it just seems like we're walking a tightrope between a generalized understanding that this is really difficult stuff and shouldn't be compared to good graphics (particularly not in the sense of a trade off) and wanting reasonable refinements of what they have already done.
I didn't post to try and shut down this argument, because it's a good one (from many of you). But a lot of people -- not necessarily in this thread -- go overboard and demand better AI and offer up graphic quality as something they're willing to sacrifice for it, and that just says to any amateur developer (to say nothing of a professional one) that you don't 'get' how their process works.
And if you want to influence their process, the most helpful thing you can do is to have a thorough understanding of it. Which means not equating an occasionally suspect AI with CA being disinterested in AI.
The troublesome thing for us is that it's difficult to test properly. You need custom battles that play out exactly the same to get reliable data. Good luck in trying that with anything more than 1 unit vs 1 unit. So we report the AI decisions on a case by case basis, which cannot usually be repeated, since battle replays cannot be saved in the campaign game, assuming we're playing single player.Quote:
Which means not equating an occasionally suspect AI with CA being disinterested in AI.
But some tests have been done and they have led to some improvements in the skirmish AI. Phalanx AI and army level AI is still lacking though, and we somehow do not have ways to provice CA with meaningful data...
patient in any case until MTW2 is released... what can i dO?
I missed it by about a week, but that was a superb post Aquitaine. Very well put and very real.
AI cards, that would be exceedlingly sweet :)