How "Starter Marriage" Thinks

[mns_95125]

I'm a little surprised you use a weighted random traversal instead of a deterministic one. Why not just BFS or DFS the option graph?

[Daphne]

I'm a little surprised you use a weighted random traversal instead of a deterministic one. Why not just BFS or DFS the option graph?

Well, consider the size of the graph. To get from the first scene to the transition to the bedroom takes 80 to 130 choices. Each scene has an average of about 2.1 exits, so we're talking about 2.1^80 to 2.1^130 possible paths. It takes the script between 0.5 to 1.1 seconds to process each playthrough.

Of course, the game doesn't have that many distinct scenes, but since the text and choices (and sometimes image) vary with each path, each scene is basically a new node depending on the path that was used to get there.

(In fact, I initially tried a DFS, and then realized that the universe would collapse into heat death before it finished.)

[ozymandias]

Well, consider the size of the graph. To get from the first scene to the transition to the bedroom takes 80 to 130 choices. Each scene has an average of about 2.1 exits, so we're talking about 2.1^80 to 2.1^130 possible paths. It takes the script between 0.5 to 1.1 seconds to process each playthrough.

I don't think the size of the graph is 2.1^130 (though that's an interesting bound), because every decision doesn't modify the *next* node in the graph; it only modifies the state vector "a bit", right? So calculating the graph is actually the interesting problem--what are the probabilities of each scene following a given scene? And do those match your desired probabilities? (And also: really? 1.1 seconds to traverse the graph once? Are you making http requests or just processing the files? Can you load the whole thing into memory and just traverse the graph in memory?) Drawing the state machine would be hard, but given that you have a way to traverse the thing pretty quickly, you should be able to build up a probability map of each scene's entries and exits. Looking for near-zeros should tell you if you have a problem or not.

It's an interesting problem, though.

[twobears]

You might consider add a counter to the different pages on the actual web site.

This would only give a retrospective view of what's been deployed (and is subject to bias by the end users preferences, but it might prove helpful)

[jwik01]

I feel quite low-tech with my paper and pencil but I'm from the humanities so

[Daphne]

really? 1.1 seconds to traverse the graph once?

Well, keep in mind that we're not just "traversing the graph." At each scene, we need to:

1. Read the state in from disk.
2. Apply any state changes to it.
3. Render the template for the page, including choices.
4. If the image is generated rather than static, generate the image (spoiler alert!).
5. Write the state back out to disk.

We could skip #1, I guess, but we can't skip #5, because having the trace of moves and state is very helpful for debugging. We can't skip #3, because that's essential for debugging the page text (and templates can have errors that only show up in particular states).

One of the other functions of the playthrough 'bot is to get test coverage on the code as it is run in the real-life system, so I try to minimize the number of special magical things that the playthrough 'bot does that the regular website does not.

Yes, there are potential optimizations to how the state is handled besides reading the whole thing in and writing it out, but that would require some pretty smart code, and the possibility for very hard-to-find bugs there is pretty high.

[afurydarkly]

And don't skip #4 cause generative boobs...