Random Game Development Thoughts

[Zed-F]

I was taking a look through some of the Civ4 Epics commentary (I haven't played very much Civ4 at all but occasionally get curious) and a thought crossed my mind.

It seems almost a given these days that good games have lots of features, and better games have more. The same applies to content. More playable races. More diplomatic options. More different kinds of units. More customizable or modable content. More sophisticated AI. More development time and dollars spent to get a bigger payout at the end.

Whatever happened to just doing a few things, but doing them really well?

Building a really good AI is really hard. It's time and labour-intensive, and difficult to justify when it often doesn't translate well into revenue, especially once you get past 'good enough that you don't get slammed for lousy AI' and start approaching 'praise for really good AI.' If I look at a game like Civ4, it has lots of strengths to its AI, and lots of weak spots, simply by virtue of how much stuff it has to do.

Let me compare that with MOO1. On the surface, there's no comparison. The Civ4 AI is vastly more capable and more intelligent than the MOO1 AI. The Civ4 AI approaches warfare with some idea of what its objectives should be, and some idea of how to acheive its goals, even if not nearly as solidly as a human. It has some idea of how to manage its cities, even if the human player can do it a bit better. The MOO1 AI? In wartime, when it mounts an offensive, it just throws whatever it happens to have built your way, showing little understanding of how to build effective fleets or which worlds are important to attack, and in peacetime, empire management happens pretty much by default.

So why is MOO1 considered such a classic even if its AI wasn't really any great shakes once you sit down and look closely at it? Maybe one reason is that the game design supported the AI rather than exposing it. The MOO1 AI's version of empire management was pretty underdeveloped, but because empire management in MOO1 is pretty straightforward, it didn't need to be terribly sophisticated. MOO1's combat AI is not very bright, but because planetary defenses are a powerful option, and because combat is nodal (only at planets, not in interstellar space) rather than spread across a map, even if the AI is foolish with its fleet, its fixed defenses can still command respect -- and the fixed defenses are not customizable and thus not subject to being screwed up by the AI.

I think it would be refreshing to see a modern game that, instead of requiring a hugely sophisticated AI, was designed with the intent of minimizing the number of key decisions the AI were required to make. It's hard to get more than one AI designer working on any given title anyway, so why not design the game so that one AI designer is enough? Sure, you get fewer options that way, but you also get less unintended emergent behaviour as well.

If there are significant areas of the game where the AI is hard-coded, that means there are significant areas of the game that can be more readily insulated from the consequences of sub-optimal decisions. If the AI doesn't get to make choices about how much fixed defenses it uses, or how they are deployed, and yet those defenses are set at a level to provide a credible challenge for an attacking player, then the AI is more free to get creative about how it designs and uses its offense. A good AI decision there can provide a nice challenge for the player, but a bad decision won't be totally crippling.

Certain game designs will tend to support this style of AI design, whereas others don't. More and more we seem to be seeing open-ended game designs that require ubiquitous AI decision-making, however. Is this really where we want games to be going?

[Kylearan]

Hi,

Whatever happened to just doing a few things, but doing them really well?
[...]
So why is MOO1 considered such a classic even if its AI wasn't really any great shakes once you sit down and look closely at it?

MOO1 is indeed considered to be a classic, but I wonder how it would sell today? Assuming a game like MOO1 with state-of-the-art GFX/SFX but similarly "simple" gameplay would be published today, I don't think it would sell very well.

I think people's expectations have risen considerably over time, and I think their expectations are closely tied to the increasing GFX/SFX standards. They see animated 3D GUIs with famous actors doing the voice-overs, rapidly approaching movie standards, and so they expect gameplay to be similarly complex. They don't see the correlation between complexity of game mechanics and AI design. What they primarily want is "more, more more!".

I've read through some of the general discussion over at Civfanatics about the Warlords expansion for Civ 4, and found several people complaining that the expansion has not introduced more resources, even more units, more wonders, etc. These were sometimes the same people that complained in another thread that Firaxis had failed to improve the "stupid AI" significantly...

If a game like MOO1 were published today, people would compare it to games like CIV and would notice at once that it would have less races/traits/weapons/resources/whatever than CIV, and simply wouldn't buy it. So in my opinion, the comparison between modern games and old classics is a bit unfair.


That leads to the question, why are people looking only at the surface of a new game, counting number of features, and do not look deeper? Because the media heavily supports this kind of thinking, IMO. I've long stopped reading computer game magazines because of this phenomenon. When a new game comes out which interests me, I read a lot about it on the net - and 98% of all sources just praise how the game makes use of that brand-new shiny 3D engine, which famous actors had been hired for the voice-overs, how many different units the game features ("ICBM explosions look GREAT!") and so on. It's almost impossible to find something about an AI's quality! No, that's not right: Lots of reports mention just how amazing it is how many man-years had been invested into this AI, and that it's better than any AI seen before, and how it adapts to the player's actions (I don't know how often I've read that statement now over the past years...). The reviews almost never provide a detailed look on an AI, listing some of his strengths and weaknesses - probably because most of the time, they simply reiterate what the publishers had told them.

It's VERY frustrating for me as a consumer to get qualified information about these kinds of things: AI, inner game mechanics, economic model used in the game etc., and today I'm relying only on RB and the LL for more qualified information. Now, if you were a game designer, would YOU try to publish a less complex game, which has game mechanics mainly geared towards supporting the AI, resulting in a less compless game world? I know I wouldn't, unless I'd be a millionaire not depending on how well my game will sell.

-Kylearan

[zilchy]

Take a look through the MOD's available for CIV and it's clear that as Kylearan say "more stuff" is what most people want and not necessarily quality. It seems to me that the most popular mods give more units, civics, nations etc... The SDK has been out a fair while and I have not seen any large AI changes even though its there for anyone to have a go at.
I don't know that it's just that the requirement is not there or that it's just too hard to write an AI for something like CIV (or any other game). It's strange in that loads of people moan about AI's quirks but as I think Sirian said elsewhere the reward for writing a good one is not there in the market place.
Also at the basic level what is a good AI? What do people want from an AI? It's clear that people don't want to be beaten everytime by an AI. Just look at any Chess program out there where most engines will beat most people so there are options to dumb them down. Yet, I've never played a chess program where the dumbing down has felt at all natural or satisfying.
I don't know what will change developers, and I'm not sure that they will anytime soon.

[Sirian]

Now, if you were a game designer, would YOU try to publish a less complex game, which has game mechanics mainly geared towards supporting the AI, resulting in a less compless game world? I know I wouldn't, unless I'd be a millionaire not depending on how well my game will sell.

I agree with Zed more than with Kylearan, to this point in the thread, although both have valid points. Letting the hype side of the industry (marketing) make design decisions en masse is a road to disintegration for the industry on the whole.

The right mix of simplicity and depth can go a long way. MOO1 is classic because it's OLD, and standards really were lower then. I'm afraid that the shine finally wore off for me when I saw the game with 2004's eyes. There are a lot of flaws I put up with back then simply because it was the best total package at the time, but which do not meet my standards today, no matter how much I WISH they could. We all outgrow our old video games.

Who would go back to Civ1 with its all-units-in-a-stack-die mechanic? Or its battleships attack settlers mechanic? Yet there are brilliant things in Civ1 (which, frankly, as far as the AI needs are as complex as Civ4's needs in most cases) that still ring true, and which make games like Civ3 and Civ4 successful. Something old, something new, something hard, something true.

Yet Zed's point remains valid. Most of what is WRONG WITH today's games arise from too much of a good thing. Like, hey, Survivor was COOL, and so was American Idol, but Reality Show Clone #117, "We hope you want more of the same, only suckier" is really way too much.

Diablo2's faster run speed may have met the market demand, but dang did it make the game easier or what? All those fancy skills on the melee classes were cool, until you realize that the only way left to play a melee char is with RANGED skills, because small tactical movements are impossible, and everything is playing too fast for all but the best mouse clickers.

MOO2 and MOO3 both stank because they piled on junk gameplay. Someone put the "realism boys" in charge and they made steaming piles of goo. More realistic goo, sure, but not much fun to play.


Titan Quest has been enjoyable for me because it took some of the best of Diablo 2 gameplay but also brought back some Diablo 1 things, and did it all in a fresh, creative new setting. The only reason I haven't been recommending that game yet is because it's godly unstable on some systems, including mine, including many (if not most) higher end systems. (Yes, it's the HIGHER END systems that are having trouble! Go figure.) I'm waiting for them to sort that out with a patch. If/when that happens, I'll talk about the game.

Really, though, any great game is something simply you enjoy doing over and over again. The more complexity they pile on, the more likely that most of the extra options will be unbalanced and therefore unfun.

Less is more!


- Sirian

[Zed-F]

I certainly agree that there are market dynamics involved -- that there are dollar-based reasons why games are built the way they are these days (and also some reasons that are based in perception and not actually tied to the market.) Game developers need to cater to the lowest common denominator in their target market, and they need to make that target market as broad as they can afford to while still maintaining a coherent game idea. If they can throw in some extra stuff to challenge the 'power gamer', then great, and most developers will make attempts to do so.

But all this stuff makes it hard to balance a game. For a well-designed game, the developers will have a vision of what the game should be. Necessarily, the more options that are included, the more diluted that vision becomes. It is possible to build a really strong AI that can challenge the best human intelligence, but only at very specific tasks. Chess AIs are a good example of a strong AI built for a tightly-bounded game, and even then we're only as good as we are at developing Chess AIs because we have decades of experience at it. The more general the task, the harder it is to develop an AI that will perform well across the board.

There are a few developers out there willing to try something simpler -- take Sword of the Stars for instance. Their vision for the game is supposed to hearken back to MoO, and in some respects it does, though my experience with the demo leaves me with mixed feelings as they seem to have made some design decisions that I'm not sure I'm comfortable with. Mainly though, these are the small developers who can't afford to tackle really grandiose projects. However, those small developers are also the ones who can't afford to ignore demands from their target market for 'more, better,' stuff... so we'll just have to wait and see if any of them actually have a tight vision for a game that they have a hope of building a solid AI for, and then stick to it.

[Kylearan]

Hi,

Also at the basic level what is a good AI? What do people want from an AI? It's clear that people don't want to be beaten everytime by an AI.

You're right, I don't want to get beaten by the AI all the time. To me, a good AI for a game like CIV or MoO would be an AI that does not cheat and still presents a challenge. I'm still quite impressed what Firaxis did with the CIV AI, but once you've grasped the game mechanics, it still needs massive boni to be competitve. I find it hard to appreciate all the improvements Firaxis has made to the AI like assembling larger stacks, making surprise landings etc. when it still needs massive production/research boni, maintenance discounts and free starting units/techs to pose a threat to the human player. If I lose, I don't feel like I've been outsmarted, only outnumbered. A good AI to me would be one that outsmarts me from time to time!

I've programmed an AI for a card game (which you can find here). It's not hard-coded, but has been evolved using genetic algorithms over several months of computing, playing countless games against itself and learning from it. After it had converged to a stable AI, it surprised me by successfully using tricks against me I've never seen before! Of course, San Juan is a very simple game compared to CIV, but this is what I'd like to see from an AI: Surprise moves, not being predictable.

Of course after hundreds of games against my AIs, I know their strengths and weaknesses as well and win most of the time, but not always. Whereas with CIV, I've won all of my first games, including my first one ever on emperor. This was mainly because I knew the way the AI plays back from Civ 3, and even in the new version it's extremely predictable. Where it should have a variety of tactics and little tricks to choose from, randomly if need be, there's only one algorithm that plays the same all the time.

I realize that I won't see such an AI anytime soon, but then I'm not a cutsomer who is not easy to satisfy.

-Kylearan

[Majromax]

Firstly, MOO3 sucked because they seemingly never finished the game. The AI wasn't just bad, it was incomplete.

Secondly, humans are good at 4X games because we group things. We form fleets/armies of units with specific tasks in mind, we create "provinces" of cities dedicated to the same or similar tasks, and we more completely divide up the opponents into blocs of allies and enemies.

With this, the human and AI aren't playing the same game anymore. The human's playing a game of a few dozen "pieces", each of which is very strong and flexible. The AI is playing a game of a few thousand pieces. Even where the AI groups units (such as an attack stack), we've seen that their groups are far less flexible than an equivalent group that a human would make. For example, in Civ4 an AI might make a "city attack" stack -- that would then proceed to completely ignore the player's incoming attack on a core city.

Until this problem is fixed or a 4X game is redesigned to enforce groupings in code (so that the palyer and AI are both dealing with the same sorts of groupings), a 4X AI won't be able to compete. I think that MOO3 was trying the latter, with its task force system and (in development) penalties for micromanagement, but the latter was unfun and the former never actually had competence on the AI side of things.

[Green_Gloom]

Titan Quest has been enjoyable for me because it took some of the best of Diablo 2 gameplay but also brought back some Diablo 1 things, and did it all in a fresh, creative new setting. The only reason I haven't been recommending that game yet is because it's godly unstable on some systems, including mine, including many (if not most) higher end systems. (Yes, it's the HIGHER END systems that are having trouble! Go figure.) I'm waiting for them to sort that out with a patch. If/when that happens, I'll talk about the game.

You may want to check ot next weeks patch.

[uberfish]

I've programmed an AI for a card game (which you can find here). It's not hard-coded, but has been evolved using genetic algorithms over several months of computing, playing countless games against itself and learning from it. After it had converged to a stable AI, it surprised me by successfully using tricks against me I've never seen before! Of course, San Juan is a very simple game compared to CIV, but this is what I'd like to see from an AI: Surprise moves, not being predictable.

My compliments on the AI. I haven't played this card game before, but it seems to do a good job of mixing up its strategies. (Observing my own learning process and figuring out the game rules and interactions between the cards was pretty fun too)

[Sirian]

You may want to check ot next weeks patch.

I'm glad it's coming soon.

I've been experiencing a reduction in crashing at lower graphical settings in combination with "holding still" when framerates plunge. Does not stop all crashing, but in combo with endless manual backups, I've been able to play and enjoy some TQ.

Would be more fun if fully stable, though. We'll see what the patch holds.


- Sirian