Maze Generation Algorithms - An Exploration

This webpage is dedicated to my exploration of maze generation algorithms .

To begin, I did some research on maze generation. There's a nice wikipedia page, which is linked above, that outlines a few of the more common ones. There's also a fantastic resource I found that has animated implementations of every "perfect" maze generation algorithm - that is, every algorithm that generates a maze with exactly one solution. I primarily used that resource for a list of every algorithm that I will be implementing, but it's also quite nice to visualize the manner in which they work.

On this page, I will be attempting to do something similar to what was done on the page I linked: animating each algorithm and briefly describing what they do. Ideally, this will become a good resource for people interested in entry-level procedural generation as well as an interesting read for those who are curious but lack a background in computer science (or graph theory, for that matter).

And finally, without further ado... we begin.

The Backtracking Algorithm

Some people also call this the randomized depth-first search or the modified depth-first search. The idea is relatively simple: The algorithm branches out in a random direction until it branches itself into a corner. Then, it backtracks, or retraces its steps backwards, until it reaches a node from which it can once again branch out. It continues this process until it cannot branch from any nodes, at which point the algorithm terminates because, by definition, the maze must be completed.

It's probably easiest to understand when it can be visualized, so here's an animation:

You can watch the concepts manifest themselves in the animation - the algorithm branches out randomly, periodically exploring into a dead-end or corner. Each node it explores to is added to the list of open nodes, and is marked in pink. When a point is reached at which the branch can no longer continue exploring, the algorithm backtracks. When it backtracks through a node that it cannot explore from, it removes that node from the list of open nodes, and the location turns white. A wave of pink becoming white is therefore indicative of a long string of backtracking.

Of course, the simplicity and speed of this algorithm does come with a few slight disadvantages. Like most maze generation algorithms, this one does NOT generate a uniform spanning tree. What does that mean, you ask? Well, "spanning tree" is just a pretentious way of saying maze in graph theory language. The word uniform, here, just means "chosen from all possible options, weighted equally." So "uniform spanning tree" just means a maze generated in such a way that it was randomly selected from a list of every possible maze to be generated. This algorithm does not do that.

Despite its shortcomings, however, the ease of implementation of thebacktracking algorithm combined with the adequacy of the mazes it generates makes it ideal as a goto algorithm for this purpose. It's not great, but it's absolutely good enough.

The Randomized Prim's Algorithm

Prim's algorithm is an algorithm designed to find a minimum spanning tree for a weighted graph. That probably sounds like a lot more fancy graph theory language, and it absolutely is. Frankly, you don't need to know what that means to understand how the algorithm works. It's much the same as the backtracking algorithm, with one key difference. Each turn, instead of continuing to branch of the "active" branch, as with the backtracking algorithm, there is no active branch at all. Rather, every single time the maze branches off, it does so from a node that is randomly selected from the open set of all nodes.

I'm dismayed to see that what I just said is creeping rapidly into the area of incomprehensibility, especially to someone who hasn't done reading on maze generation, graph theory, or something else related. What "randomly selected from the open set of all nodes" means is this: Remember all those pink squares in the backtracking algorithm? Every once in a while a wave of them would disappear in quick succession. Well, in the randomized Prim's algorithm, every single branch-off comes from a random pink square. Instead of long winding branches weaving themselves through empty space, the maze appears to spread like a virus from its source. Have a look:

Once again, you can watch the algorithm do its thing. The active nodes, which are shown in pink, are usually towards the ends of the maze. This is because while there are inner nodes, part of the open set, that cannot be expanded from, if at any point those nodes are selected from the set of open nodes and discovered to be un-expandable, they are subsequently removed from that set (and marked in white). So an inner node marked in pink will usually turn white relatively quickly because once its branchability is tested and determined, it is marked accordingly.

Prim's algorithm produces maze that are not quite as aesthetically pleasing as the backtracking mazes, and like the backtracking algorithm, the randomized Prim's algorithm does not generate a uniform spanning tree. For these reasons, I consider Prim's algorithm inferior to the backtracking algorithm. It was fun to implement, but its practicality is somewhat lacking in my opinion.

Eller's Algorithm

Eller's algorithm is hard. Really hard. It uses sets to expand different paths of the maze, one row at a time. The way in which it works is quite complicated, but I'll do my best to explain it briefly in case you're curious. If you're not especially technical, you might want to just take a look at the animation and then skip to the next algorithm. Here's Eller's algorithm in action:

It works like this. At the beginning, it looks at the first row of cells. Each cell is independent, and each set of cells contains just one cell. Then, it randomly joins cells (and the sets they're a part of) so that there are less sets and some sets have multiple cells in them.

The next thing it does is create downard connections. For each cell in the row being looked at, it randomly creates or doesn't create a connection down to the cell below it. If the connection is created, the lower cell is added to the set of which the upper cell is a member. Otherwise, the lower cell is added to its own new set, containing just it. While this is happening, the algorithm also ensures that every single set from the upper row has at least one continuation down to the lower row. This ensures that every single set currently stored has at least one cell that is a member of it on the bottom row.

Through this method, along with a little set and graph theory magic that I won't even bother to try to explain, a perfect maze is created. It's one of the coolest algorithms I'm going to implement, as well as one of the most difficult. But I think it works like a charm.

The Randomized Kruskal's Algorithm

This algorithm is actually really cool. It's similar to the randomized Prim's algorithm in that the unaltered version is used to generate minimum spanning trees of weighted graphs. The randomized version randomly removes walls until the maze is completed, and I think among the algorithms I've implemented so far, it's among the coolest and most interesting to watch. Have a look:

The way it works is also surprisingly simple: First, it initializes the maze. Each unconnected cell is a member of its own set (sound familiar?) and each wall dividing two cells is a member of the same set of all walls. Next, walls are selected at random and removed if and only if the two cells it divides are not a member of the same set. If they are of the same set, the wall is not removed but it is thrown out of the set of all walls, out of consideration. It then joins the two sets of which the cells are members. It continues doing that until there are no walls left in the set of all walls, at which point the algorithm terminates.

This one (along with Eller's algorithm) is my favorite to watch because of the mesmerizing fashion by which walls are removed. It's also pretty interesting that at the end of the algorithm, every single cell in the maze is a member of a single set. I never thought about mazes that way, but any perfect maze must be one set (so every cell can be reached) and each cell must have only one connection to every other cell for it to be considered a "perfect" maze - which explains why cells are only joined if their respective sets are different. Only one path between previously disconnected sets means only one path from one cell to another. It's pretty cool.

Enough of my nerdy ramblings, though. On to the next algorithm.

The Aldous-Broder Algorithm

The Aldous-Broder algorithm is an algorithm for generating uniform spanning trees of a graph. Remember the definition we gave for those earlier? Uniform Spanning Tree means "a maze generated in such a way that it was randomly selected from a list of every possible maze to be generated." But I'm going to simplify that definition even more by saying this: A maze generation algorithm that generates a uniform spanning tree is the only maze generation algorithm that could, theoretically, generate every maze possible. Hopefully, that makes a little more sense. They're probabilistically superior to most other maze generation algorithms, but that property makes these algorithms much more inefficient than the standard ones.

This is also one of the few algorithms that is best explained before, not after, the animation is presented. It's almost trivial in its simplicity, and its implementation was not difficult. It starts by picking a random cell in the maze. Then, it jumps to a random neighbor of that cell. If that cell has not yet been visited, it connects it to the previous cell and marks it as visited. Then, it jumps to another random neighbor and does the same thing, continuously jumping randomly until all cells have been visited, at which point the algorithm terminates. Like I said, really simple.

This is what it looks like:

Like I said, the algorithm is very inefficient, but it does do the whole uniform spanning tree thing, which is really cool and also unique. There is only one other maze generation algorithm that generates uniform spanning trees. Let's take a look.

Wilson's Algorithm

Wilson's algorithm uses loop-erased random walks to generate a uniform spanning tree for a graph. I'll explain what that means, but for this algorithm I think it's best to present the animation first. I'll warn you that it can take a while to converge on the first step, so you can click on the bold letters below to change the size of the maze.


Now to explain that "loop-erased random walks" thing. Basically, the way this algorithm works is as follows: First, it picks a random cell (square) and turns it white, marking it as part of the maze. Then, it picks another random cell and starts looking for a path between the two cells. When that path wraps around and runs into itself, a loop is created. That loop is erased, removed from the path, which then continues from the base of that loop (which is now gone). That's the "loop-erased random walks" bit, and that's why at times the first step takes a while to complete - it's trying to connect a random walk to that one square that we picked in step one.

Once the first loop-erased random walk reaches that one lone square, the entire path becmoes part of the maze. Then, we go back to step two, which was to pick another random square and start a random walk from there. This time, however, instead of having to reach a single square, this walk just has to reach any part of the existing maze, which is now much bigger than a single square. Once that random walk reaches the maze, another is created, and another once that one completes. This occurs continuously until the whole maze is completed, and this technique, just like the less efficient Aldous-Broder method, creates a uniform spanning tree. That's right, this bizzare and unconventional algorithm can create any maze possible. Isn't that cool?

Perhaps the most peculiar and difficult-to-understand aspect of this algorithm is the property that regardless of the method used to select new cells from which to begin the loop-erased random walks, a uniform spanning tree is created. Phrased for those with a background in CS, this means that whether the origin cells are selected iteratively, randomly, using alternating source iteration, or any other convievable method is irrelevant to the functionality of the algorithm. Phrased for a curious reader without an understanding of what that means, one could instead say this: If the squares from which the random pink paths begin are not selected randomly, the algorithm still works exactly the same way, and it can still produce any maze possible. Pretty crazy, huh?

For example, on this page, if you scroll down a bit, there's an implementation of Wilson's algorithm that selects origin cells iteratively on the grid - that is, whenever a path is closed and made part of the maze, instead of picking a random square to start the next path, the next square is selected, where "next" just means "first, starting from the bottom left." You can watch it work, and in some ways the behaviour is very different from mine. But it still works, which is one of the coolest things about this algorithm.

But enough nerding out - this is supposed to be accessible to everyone, after all. On to algorithm number seven.

The Recursive Division Algorithm

This one isn't that great. I included it because I'm including every one in existence, but this one is really nothing special. It divides the maze into two sections, with one path in between. Then, for each of the two sections, it treats them as a maze in and of themselves. It cuts each of those into two, then cuts each of the fourths in two, and so on until every compartment is too small to further divide. Simple and unimpressive, not to mention the fact that the mazes it generates are very obviously made using this algorithm - there are long, unbroken walls that are characteristic of the divisions. But regardless of its shortcomings, I did animate it.

The Hunt and Kill Algorithm

The hunt-and-kill algorithm is in many ways the same as the backtracking algorithm in that it creates long winding paths. However, where the backtracking algorithm backtracks, the hunt-and-kill algorithm simply searches iteratively. This means that when it curls itself into a corner, it gives up on that and looks for the "first" (left to right, one row at a time) cell that neighbors part of the maze. It then continues building the maze from there. Take a gander:

This one's interesting. It branches out randomly, then it iteratively fills in the holes it creates through that random branching out. The end result is an animation that creates a few crazy passages and then fills holes in a pattern that resembles how one reads text on a page - left to right, then top to bottom.

Full disclosure, there's not a whole lot to say on this algortihm, since it's so simple. Let's just move on.

The Sidewinder Algorithm

This one's kind of cool in that it's the only other algorithm - besides Eller's - that works on a line-by-line basis. It uses even less memory than its cousin, and it's slightly faster, but at a cost. The mazes don't have a noticable bias, but they're not especially aesthetically pleasing. They have lots of very short dead ends, and there's the obvious shortcoming of, you know, the giant unbroken path across the top. But the way it works is still kind of cool to watch:

The way this one works is pretty surprising in its simplicity. First, it creates a long bar across the top for reasons that will become clear momentarily. Next, it goes to the second row, and that's where the fun begins. It starts on the first cell, and randomly decides whether or not to continue each time it goes to the next cell. If it continues, it adds that cell to the current set and joins it to the previous cell. If not, it keeps the wall between them and empties the current set, adding the new cell to it. Finally, if the current set was just emptied, it picks a random cell from the previous set and carves upward from there into the row before. That's why the top row is a continuous path - since you can't carve a path upward, it has to continue every time until the whole row is one path.

It's not quite as sophistocated as Eller's algorithm, and that simplicity can almost be seen in the apparent inferiority of the mazes it creates, but this algorithm is nonetheless cool. It's even faster then Eller's, uses even less memory, and, like Eller's, can create infinitely tall mazes. It's pretty cool.

Now, believe it or not, we're on to our very last algorithm. Some people like to save the best for last, but that's no fun, bcause then you have to wait for the best, so this algorithm is actually the simplest (and arguably the worst) of all the algorithms we cover. It's called...

The Binary Tree Algorithm

This one is irksomely, frustratingly, infuriatingly simple. You litearlly go, one cell at a time, iteratively, and randomly carve a passage either down or to the right. If a passage can't be carved down, it goes to the right. If a passage can't be carved to the right, it goes down. And if neither is possible, well then I guess we're finished.

Like I said, infuriatingly simple. The mazes it creates are also ugly and have a diagonal bias, which doesn't exactly help. And it doesn't work one row at a time, like sidewinder or Eller's. It actually works one cell at a time, which is even stupider. Basically, this algorithm is dumb.

But wait. Is it? Sure, it creates miserable, sorry excuses for mazes, but they're still technically perfect mazes, and that fact alone makes the simplicity of the algorithm at least a little bit cooler. I mean think about it - the fact that such a trivial set of rules comes together to create a spanning tree (a perfect maze) is kind of interesting. The algorithm itself is the worst, but it's not the least cool. I would award that title to the recursive division maze, which is kind of underwhelming both in its functionality and its final products. But enough talk. Before we wrap up, let's take a look at this pathetic, wonderful algorithm:

It's like... kinda cool to watch, I guess? Whatever. It's there, and its presence means that we've finished. We've implemented all 10 maze generation algorithms.

But Like... Why?

Frankly, that's a really good question. Why did I do this? What purpose is there for not only implementing, but animating every single algorithm in existence that does exactly the same thing? The answer is pretty simple: More than anything, I love to teach. I love instilling passion and tapping into curiosity, and teaching does exactly that. I don't expect that anyone who reads this page will become inspired and develop a firey passion for the world of computer science. I do hope that after reading this page, or only skimming part of it, or even just watching a few of the animations, you have a slightly increased understanding of how these algorithms work. Watching them in progress is the next best thing, short of actually implementing them in code, and the latter requires experience and a whole lot more time than the former. That's why I did this. To educate.

Well, that, and it was colossal fun. It was a great learning experience for me, too, and my first passion is learning; education is a byproduct of that. I loved doing this and the knowledge I gained throughout the process. It's one of my proudest achievements in programming, aside from perhaps my 2x2 Rubik's Cube solver in C, a project that is not yet on Github because I used it for my AP Computer Science Principles class and I can't have it uploaded just yet. But this one, if it doesn't exceed that in difficulty, certainly approaches it. And I think this one was even more fun, which makes it better. I had a blast making this webpage, and I hope you enjoyed reading it.

As always, the source code is available on Github for those of you who wish to skim or peruse it. My other work can be found there as well, and you can contact me at any point with questions, suggestions, or salutations. This project was finished on the same day I graduated high school, and my next project (much easier) will be to implement A* to Solve these wonderful mazes. Then perhaps I'll animate that, so people can learn a bit about graph traversal. But for now, I'm done for a few months. I'm off to a camping and hiking program for June, July, and part of August to cater to my love for the outdoors. So hopefully sometime in August you'll hear from me again. Thanks for reading!