Good procedural generation requires a lot of effort to make work. The quality of the output? While in theory, can be as high as a handcrafted level - A random level generated could, with the right algorithms, match a human crafted level, and there are likely other options the computer could spit out that are as enjoyable to play as to the human-made level while being distinct from anything a human might come up with. In reality, however, they're simply not that great at generating the superb levels that some games have the occasional example, or even the very good levels a step or two below that which most good games have, but instead at best tend to generate billions of alright levels. This is easy to visualize if you simplify the scope to something as simple as Sudoku. On paper, any human crafted grid, or series of logical steps deliberately inserted by a human, could have been spat out by a computer. But a computer is unlikely to generate a puzzle that starts in a satisfying shape without specific prompting or has a weird series of logic to generate an amusing chain of pairs that unravel at the end of the puzzle enjoyably. Or tell what logical steps are enjoyable to make rather than coming across as complete codswallop to a solver. Fundamentally, however clever you make the algorithm to solve these issues, a computer is fundamentally unlikely to generate a memorable puzzle. While that's not an easy task for a human, any more than writing a great novel is, it is at least something human setters can aim for. So why put the effort in if it's not going to generate as entertaining levels? Given the choice between infinite alright or better levels, and 100 very good or better levels, which would you choose to play? If you knew that would be the output of your labour, which would you choose to make? Because for me that answer depends on what the game is trying to do. Permadeath sucks if you always play the same levels - Even in older titles that lacked save points. You lost on level 10, go back to start. It's not so bad when each time you play you'll get a different first ten levels. No matter how good a level is on the first play, it's rarely as enjoyable on the 100th play - Even if it's better on the 10th. Familiarity can breed contempt, even in games designed around revisiting the same areas over and over. Permadeath is a very good example of why you might need procedural generation, but there are a couple of other cases - games with a fixed world but having a scope requiring far more content than what a human team could create, for example. In other games it doesn't matter if the procedurally generated environment isn't interesting in and of itself, because the game is about the impact you have on that environment - Minecraft is the obvious example here, with a procedurally generated world that you modify. In those examples, the procedural generation isn't so much content as a blank canvas and writing prompt all in one, and in those cases the weaknesses of procedural generation stop being particularly weak. Similarly, the tabletop game How To Host A Dungeon gives the player an algorithm to run manually by rolling dice to generate features, build fictional histories and landscapes on the terrain they're drawing as they play based on this process. Procedural generation gets an undeserved bad rap in some circles, and others are overly enthusiastic about it. It's a tool, useful for some situations, inappropriate to use in others, and which can be used badly even in those cases where it was appropriate to use.
- Stephen