Slashdot Mirror
IEEE Introduces Mario Level-Generation Competition
bgweber writes "Last year, the IEEE conference on Computational Intelligence and Games hosted a competition to determine who could write the best AI for playing Mario levels (YouTube video). This year, the conference has expanded the competition to include a track on level generation as well, where the goal is to generate new levels online procedurally. Submitting an entry is as easy as implementing a Java interface that performs procedural content generation. The implications of this competition are techniques for greatly increasing the replayability of games, since each gameplay session could present new levels to the player."
Congratulations adventurer! Your quest is at an end for you have reached the home of NetHack.
I fear the day when the singularity occurs and we peons are forced to play machine generated levels like this this for the amusement of our robotic overlords.
'We are trying to prove ourselves wrong as quickly as possible, because only in that way can we find progress.' RPF
I'm glad it's not a Mario-Level generation competition. No one wants to power their circuits with a generator designed by a cartoon plumber.
It would probably just be a gorilla and a dragon forced to turn a mill wheel that creates electricity by grinding up star bits.
I get the point of the competition... but I seriously hope this won't actually end up being a feature of future versions of mario. Part of the whole point of mario is learning where all the secret bits are and showing off to your friends who aren't in the know. This would also defeat the purpose of retrying the hard bits until you finally conquered them and make the game incredxible unrewarding
As a non-programmer, this statement is a little intimidating:
"Submitting an entry is as easy as implementing a Java interface that performs procedural content generation."
WTF am I supposed to do with this? I only clicked to this post because I like mario.
...::----::...
I am in no way affiliated with this sig.
This looks like something I could get into instead of creating a new game every week out of boredom (and unemployment), but why java? I know C#, which is basically java anyway, but still...
Randomly generated dungeons are fine and generally fun because the entire point is to explore. Randomly generated Mario levels are going to be more frustrating than anything else because the AI is going to have very little knowledge of difficulty let alone themes and re-playability.
The entire point of Mario is consistent levels with well timed jumps to reach secrets. I'm not sure if I want levels generated by computers. 50 quality levels are better than 100 AI generated levels.
Taxation is legalized theft, no more, no less.
I've been saying this for years! Random level generation in first person shooters, racers, and platformers doesn't seem terribly complex to me. Why hasn't anyone introduced this into a game yet?
To expand on it, random levels in first person shooters could also be changing during game play so that the 'round' never resets and the game continues endlessly. Any time a zone is unoccupied it could be redrawn. Or divide the map into a grid and randomly load the grid spaces with pre-designed sections, then when a section is free of players load a new section. The map will be constantly changing and the environment will be challenging. I believe this would effectively eliminate campers who memorize the best spots on maps and just hang out there the whole game.
No sig for you. YOU GET NO SIG!
Dino run http://www.pixeljam.com/dinorun/ for example is doing a great job of generating a unique level each time you play. The real challenge is to maintain the structure of a complex level that is entertaining, rather than making an endurance test (as I can imagine the Mario levels becoming) Of course, a little bit of human brain applied, users to rate the levels, and pick sections of their favourites and systems to integrate more of the good stuff/success would allow an AI to 'evolve' games and game universes.
Waiting for the other shoe to...
Tthat would be C#.
They're always talking about how FABULOUS C# is.
River Raid, the old Atari 2600 game from Activision, uses a pseudo random number generator to produce the game levels. It was a great way to avoid putting the levels in the game's limited ROM (2K?).
Old but good ideas just keep coming back.
"Luncheon meats make the sawdust in your stomach explode."
The secret to great auto-generated level is to have an AI able to determine how "fun" the level is going to be for a real player. There are tricks to help a bit (eg when the reflexes needed are outside humans reach the level won't be fun and on the contrary if the AI only needs to go in a straight line and fire once in five minutes to finish the level, the human won't it fun either).
But basically to make a level really fun, you need a *very* good AI, because you need an AI that can determine what "fun" means...
We're not entirely there yet but it's coming :)
> The implications of this competition are techniques for greatly increasing the replayability of games,
> since each gameplay session could present new levels to the player."
Utterly incorrect. People have this conceptual idea that gameplay is about merely providing a framework in which people exercise their skills. It's utterly wrong and I'll demonstrate why.
Back in the 80's, there was an air-combat game. Think it might've been F15-Strike Eagle... which included the concept of random missions in which you were sent out to hit one random air target and one random ground target for each mission.
It was the most boring thing I've ever seen. One random target is the same as another. And it very quickly becomes a case of "Why bother?". There's no progression, no reward. It's just a way of playing the same thing over and over again.
In the ensuing years, I've viewed a lot of games. And the one truism I've always found is that the length of the game and the amount of enjoyment I get out of it is directly related to the amount of information content the developers put into the game.
This is why the various Sim games bored me rigid. They have no information content. They provide a sandbox, a set of rules and let you go. To a certain extent Civilisation suffers the same problem, although the campaigns mitigate this to a degree. That's all very well if you want to play around but most of the games I enjoy playing most contain unique scenarios and ideas put forward by the developers which contribute to the information content inherent in the game.
Think of information content as the number of decisions and sets of consequences which the developers have explicitly coded for. For example, take a game like Uncharted 2. Say you have the possibility of collapsing a bridge as a gameplay goal. The game plays out with you either having collapsed the bridge or not. In the context of the story it could potentially shift between two opposite extremes, but in either case, the developers have explictly developed further decisions and consequences.
Now I know that branching pathways have a finite limit, because the development effort is effectively the sum of all the branching pathways that decisions allow. But I'd argue that a finite set of pathways is vastly preferably to a bunch of decisions which have a totally arbitrary effect on the outcome.
For example, in Civilisation, the exact placement of your home city has many potential possibilities, but to a large degree there's very few differences between them. Oh, the placement relative to resources and the coast is relevant, but on the whole it's a reaction to the randomness of the game. As such, it's exercising a skill, not giving you an opportunity to make meaningful decisions.
I've played CIV and enjoyed it, but I can't play it more than once every six months or so. It's just not interesting to me to repeat the same fundamental operations over and over again. I prefer Fallout 3 or Dragon Age. Dragon Age has extraordinarily high information content which is why it provides entertainment for so long. Fallout 3 actually has low information content relative to Dragon Age. Random encounters aside, there's just not that much to do beyond exploring or following the main narrative. And that narrative is not long. You'll find that most of your time in Fallout 3 is spent digging through minutiae in various locations, not exploring the game itself.
So the idea that you'll get extra replayability out of random generation of levels is completely false. You'll get a random experience which has no information content behind it. It'll be valueless except as a reaction test.
So the idea that you'll get extra replayability out of random generation of levels is completely false. You'll get a random experience which has no information content behind it. It'll be valueless except as a reaction test.
What if the game itself is nothing more than a reaction test? Sounds like random generation would be the way to extend playability doesn't it?
Also, saying that a game is nothing more than a reaction test is a gross oversimplification in many cases. Especially if your reactions are not limited to A or B or C, but extend from A to Z along with various modifiers, say for example ALT, CTRL, or SHIFT. Successfully navigating a decision tree where your choices are limited to A,B,C would be trivially easy. Successfully navigating a decision tree where your choices are extensive is not nearly so easy.
I personally am most impressed by games that are 'compact', in that they define as little as possible and let a world expand from there. I actually picked up the terminology from Python, which has the philosophy that a good language has a few core ideas and everything expands from that.
But anyway, so you picked a few boring games (although I disagree about Civ) you didn't like. But lots of people like games with "repetetive actions" - Tetris probably being the most famous. I think of a game like Starcraft which seems like a pretty small game, but the huge amount of intricacies people have extracted from it is incredible. Contrast that with a 'high information game' where everything feels pre-scripted. I hate the feeling of the 'invisible hand' guiding my experience. I prefer the 'here are the rules, now run with it.' To each their own.
As for random level generation, obviously on some level, there is no reason why a human should be any more capable of creating a fun level than a computer. There's also no reason why a computer shouldn't be able to write a good song. Well, there are reasons, like the human experience, but with sufficiently advanced AI we could have a computer experience that.
Instead, we will have to hope that instead of true AI, you have a human carefully define the kinds of constraints that can make a level more fun. A game like mario is a pretty well defined domain, arguably far more defined than 'rock music'. So I imagine a computer will be able to do an adequate job, although we are as far away from truly innovative randomly designed levels as we are from a truly innovative computer musician.
For those interested, the IEEE Conference on Computational Intelligence and Games is being hosted by the IT University of Copenhagen, and Mario AI isn't the only competition. There's also simulated car racing, Ms. Pac-man, 2k Bot Prize and the Starcraft RTS AI Competition (!!!).
Cryptimus has a point, this thread is pretty broad, making a 'game' is not the same as being 'entertaining' or 'successful', or whatever narrow goals you set out to do. It is all about quality and focus on genre, rhythm, speed whatever you like.
Waiting for the other shoe to...
Perhaps those who say computer generated games can't be fun should check out this thesis by Cameron Browne about Ludi.
Thesis on Automated Generation and Evaluation of Recombination Games: http://www.cameronius.com/bio/publications/thesis-2.47.zip
The video of last year's winning AI player is impressive, but it's not anything like watching a real person play. For next year, they should add a Mario Turing contest. The goal is to make an AI player that is the hardest to distinguish from a human player.
This would be relevant to game design, as having opponents and allies who seem to act human can greatly improve a game.
I've played CIV and enjoyed it, but I can't play it more than once every six months or so. It's just not interesting to me to repeat the same fundamental operations over and over again.
I like story games like GTA IV (just started ballad of gay tony) but I also play Freeciv and AlphaC compulsively. I'm not super amazing, though, for example, I just beat a game of AlphaC on Librarian level (still a couple from the hardest) in the last ten turns of the game. Incidentally, while you can play these games on maps which represent real places, the scale of the games is totally unrealistic anyway, so why bother? The randomly generated maps add very much to the game. They teach us that your starting position on the globe is very important, as if we didn't know; but it teaches it at a visceral level, which I find to be one of the primary benefits of video gaming. For example, when I had my 1989 240SX I used to obsessively use it in every racing game in which it was available, and given that street racing games were fairly ubiquitous at the time, it was not difficult. When I took it to Fort Bragg (and environs) I was making a very easy turn in wet weather, with good rubber, but some leaf slime on the road caused me to begin fishtailing (this was before I got a limited slip.) This was before I got into performance driving of the vehicle and installed race suspension and what not, but I was able to instinctively correct for the slide because I had developed reflexes. It is said that visualization is as effective as practice, when an activity is visualized correctly, so this should come as no surprise.
So the idea that you'll get extra replayability out of random generation of levels is completely false. You'll get a random experience which has no information content behind it. It'll be valueless except as a reaction test.
I think it's most applicable to competitive games. Super Smash Bros. for example. Or, taking that to a logical conclusion, levels for FPS deathmatch. One of the great frustrations of a new player is that the old hands know the maps well, and you spend a lot of time getting ambushed. Wouldn't it be nice to be able to generate deathmatch maps procedurally?
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
Shit like this annoys me to no end. You haven't demonstrated ANYTHING about truth/false, better/worse, you've just told us in MANY words what kind of game you like. Yet you keep using words as if your taste was some kind of objective truth. And as if you had done research in the area ("truism", "information content"). Chess and Go don't have a story, yet people have been enjoying it for a while. I played SimCity with my g/f, the stories wrote themselves. I played it on my own and I didn't need any story, anyway. Multiplayer shooters' stories serve only as a framework for the same thing repeated ad infinitum and they're almost the ultimate in (quick) reaction test -- yet they're among the most popular games.
Switch back to Slashdot's D1 system.
How do you programmatically calculate a number or scale of the perceived entertainment value of a Mario level? Beside being able to complete the level and the difficulty, is there something objective one can measure, or does it boil down to simple taste?
If not, then the competition is kind of pointless, since the winner would be arbitrary.
In the ensuing years, I've viewed a lot of games. And the one truism I've always found is that the length of the game and the amount of enjoyment I get out of it is directly related to the amount of information content the developers put into the game.
Maybe for you. But I found both Diablo II and Torchlight to be far, *FAR* better with its randomized dungeons than Titan Quest and given their relative sales it appears most gamers agree with me.
Some degree of randomization is also vital for games like Civilizations, you may not like it but I guarantee you there'd be a *lot* less Civ players if the game proceeded in typical jRPG fashion, with only a handful of pre-determined scenarios playing out depending on the amount of hidden 'points' you've gotten during your current playthrough.
No problem is insoluble in all conceivable circumstances.
It would be a mistake to deliver new levels each session to the gamer. Part of the reason Mario is fun is that you can practice the level, learn all of its secrets, and maximize your score and efficiency in playing it the whole way through unscathed. Subsequently demonstrating your expertise to friends and acquaintances is another layer of the fun, and then on top of that being able to discuss the intricacies of a level and reveal new techniques and details to each other is the icing on the cake.
Generating random levels, if they're fun levels, would be a great achievement. But any game that implements this would need to allow for saving of the level and sharing of the level so that the elements I list above can still be part of the overall experience.
Back in the 80's, there was an air-combat game. Think it might've been F15-Strike Eagle...which included the concept of random missions in which you were sent out to hit one random air target and one random ground target for each mission.
Back in the 90's there was EF2000, a flight sim that included a dynamic campaign that could simulate a complete war. All missions where automatically generated, but they weren't random, they matched the current state of the war. You could even see friend and foe flying around on their missions, while you are flying around on yours. The world was 'alive' and it was the most friggin amazing thing ever.
Random is boring when it has no logic behind it and thus no story to tell or play. When random on the other side means a dynamic evolving world that itself becomes the storyteller it can be the most fantastic thing ever, because the game will stop feeling like a serious of script triggers and instead feel like a real world. My problem with todays games is that they simply feel like a roller coaster ride, it looks fun and exciting, but if you have played games for a while it all just looks fake.
They have already started the judging.
Next time, PLEASE add some "SPOILER ALERT" warning. Goddammit.
The aim of the competition is not merely to create random levels, but to generate levels that are tailored to particular players' playing styles and skills. The level generator gets as inputs how well a player performed on a test level and various metrics detailing e.g. how much time the player spent jumping and running and how many fireballs were fired. The level will then be judged by the player who played the test level. See more about the rules at: http://www.marioai.org/LevelGeneration/Interface Also, please note that those who submit a level generator can also submit a paper about it to the CIG conference. http://game.itu.dk/cig2010/?page_id=483