This is the second of several posts about James Ryan’s dissertation, Curating Simulated Storyworlds. The previous post looked at chapters 1-3, which set out the concept of the dissertation and documented the pleasures of emergent narrative.
Here I read Chapter 4, concerned with the pain of emergent narrative, including critiques from other scholars and projects in emergent narrative that have failed; and Chapter 5, in which he presents his argument for curationist emergent narrative.
The major issues Ryan identifies with simulations are:
Boringness. Some simulations are simulating events that aren’t that engaging, and therefore they will never have the range to compel readers. (Something I was wondering about while reading chapters 1-3.)
Granularity extremes. The system is operating on either too large or too small a scale. As an example, Ryan showcases the system that controls how drinks may be taken in the Saga II story generation system, with an arguably excessive focus on moving objects from hand to hand.
- As a side note: this is a granularity of state that most text adventure games wouldn’t bother with. There are some exceptions, though a few of the most granular works I know of were also never finished: for many years NK Guy worked on a game code-named Hamsterworld, which attended to player clothes and body parts (and many other systems) with great precision; of Gunther Schmidl’s And the Waves Choke the Wind, only a first few scenes were ever released. TADS 3’s library supports more in this range than any other text adventure world model I’m aware of, and handles some of the related challenges around making small actions implicit when they aren’t individually very interesting, so that at its best, the granularity of the world model becomes invisible except when there is something down in those details that really does interfere in the player’s intended action, at which point the consequence is reported. Return to Ditch Day remains one of the best examples of this kind of work, and Eric Eve’s work is also exemplary here.
Lack of modularity. The idea here is that elements of the simulation must be small and reusable; otherwise, it isn’t possible to recombine them in interesting ways. To illustrate this issue, Ryan looks at Sheldon Klein’s murder mystery generator, an example I haven’t seen written up particularly often (though perhaps I’ve been looking in the wrong places).
Lack of abstraction. Here, Ryan argues for the value of simulators that can cast different characters in different spaces and situations, rather than retelling (possibly different) stories about the same set of characters and events, since if we have a large number of stories about different characters, the appeal of the vast and the appeal of the ephemeral are preserved. (These are key features of the aesthetic of emergent narrative, as Ryan lays these out in earlier chapters.)
I am not sure what I think about this one. I will grant that the repetition of the same characters can give a kind of sameyness to story generators — though some systems, from Fallen London to Rafael Pérez y Pérez‘ Mexica, refer to characters by title or function in order to avoid the concrete effect of granting them a name.
Modeling gaps. This refers to places where it seems the simulation ought to cover some possibility or set of actions, based on what else is modeled, but for some reason certain elements are omitted.
Causality issues. Here Ryan describes how simulation causality can be too diffuse to make for good storytelling, especially in systems that rely on utility scoring where many different aspects of world state could all be considered to partially explain a particular outcome. (He gives a detailed example based on trying to interpret consequence in Prom Week, which is especially valuable here.) Though I’ve encountered this phenomenon, I haven’t seen the problem labeled or analyzed in depth before.
The solution Ryan proposes — contingent unlocking, where some events explicitly are made possibly by a finite set of prior conditions, and causal bookkeeping, where the system somewhere records how a particular outcome has been made available — will apparently come back in later chapters when he talks about his own work.
It’s a method we also used to some degree in Versu, where characters could record a string that represented why they’d adopted a particular attitude towards the player; and for that matter I use it lightly in my Choice of Games work in progress, which is not a simulation of the kind Ryan is talking about at all, but I still find it useful for the sake of later callbacks to be able to recall, say, the worst thing one character has ever done to another.
After these, Ryan next identifies pains of curation, and this is where the gloves come off.
No telling, no curation, poor curation. Since, in Ryan’s view, curation is essential, a system that has none is in a difficult state. Mere simulation traces from an engine aren’t enough to count as a narrative, and sometimes they look almost indistinguishable from a debugging console log.
During these sections, Ryan digs more into the idea of curation by backward chaining through causalities, which he brought up when discussing Labov in earlier chapters.
Poor presentation. This includes bad prose, often generated.
Failure to mount. Here Ryan drives particularly into the issue of narrative simulation systems that are presented to the public only via academic paper descriptions or in the form of sample output, instead of ever being rendered into a media experience.
Aesthetic posturing. Calling anything “aesthetic posturing” seems rather to presume some values about what’s within, but I don’t entirely disagree with what follows. Ryan here argues that generative artworks are interesting in partly because of the way they use and expose their generative nature, and that it’s not primarily interesting to come up with systems that can duplicate entirely human creations.
On page 108, Ryan gets into the question of what makes a story tellable, and suggests a process for finding narratable events that consists of finding the most interesting final event, working one’s way backward through causally linked events, and stopping when one reaches an event that requires no explanation. (Here he is drawing on some pre-existing scholarship as well, especially from William Labov.)
This kind of question relates quite a bit to a challenge I’m personally interested in: if we are building virtual humans or digital beings, then one of the things we need to do is accumulate a history for those characters, and give them the ability to narrate their own experiences. Being able to account for why you are doing something is fundamental to the construction of a persona. So one may imagine virtual beings who interact with the real world in some fashion and record events to an ever-expanding database of experiences, perhaps at the same time tracking information about the valence of those experiences, and causal links. A bad experience speaking at a conference, perhaps, might be marked with a negative valence and give the character a distaste for public speaking.
A footnote in Chapter 4 includes these sentences:
I think we can do full-fledged emergent narrative without losing the evocative aesthetics of the computational, but doing so means carrying out a procedure of curation. This is the primary call of this thesis.
This is what we get into in Chapter 5. Ryan begins by rejecting drama management as a good solution for emergent narrative, on largely aesthetic grounds:
…when a simulated storyworld is modulated through the intervention of an external system—a model of creative writing, a drama manager, a plotlevel narrative planner—it no longer works like nonfiction. By the interventionist pattern, events spawn according to the policies of a modulating system, which means they do not emerge out of the happenstance of simulation. They do not actually happen, and they do not feel like they actually happen.
I am not sure I would myself take such a hard line, but I would say that building a drama manager alongside a simulation system is often very painful. Building a simulation tends to be an expansive process in which one wants to add more and more to the expressive range of the system, and make it capable of doing more; meanwhile, a drama manager often tends to constrain what is actually presented by the system, or fiddle with it to such a degree that the expressive range is significantly narrowed again. It’s possible to spend a good bit of time in a cycle of adding a lot of simulation content, and then adding a lot of drama management to ensure that (in practical terms) that content is rarely or never seen.
For a system where drama management is a priority — and this is often the case for interactive narratives where the system needs to riff on the player’s input, and where we don’t have the luxury of running a simulation system through hundreds or thousands of frames before presenting any results to the player — I generally try to design from the dramatic production angle first and select for abstractions that will support the drama, choosing what is worth bothering to simulate purely on the basis of what sorts of dramatic outcomes I’m seeking.
I’m not sure Ryan would actually condemn that approach, but it’s trying to do something different from the emergent narrative that interests him. The drama-first approach is a route into interactive fiction rather than procedural non-fiction.
Next, Ryan considers architectures that support curation approaches, including feedforward architectures (where a simulation is used to create the material for a story that will be presented in a second medium) and feedback architectures (where the story is told within the simulation somewhere else — e.g. by an NPC telling a story about what happened within the generated world). I find the feedforward possibilities intriguing, but the feedback approach closer to what I actually need in a majority of my own projects.
In addition to the ideas of contingent unlocking and causal bookkeeping, Ryan proposes that one might use sifting patterns (matching particular formats of event in the source material) or sifting heuristics (more loosely defined) to determine which events generated by a simulation would be worthy of narration. He also does discuss the fact that some simulations are more capable than others of producing the raw material of interesting stories, which was one of my questions back in Chs. 1-3.
4 thoughts on “Curating Simulated Storyworlds (James Ryan) – Ch 4-5”
This is really interesting, to the point where I’ve been inspired to download the full thesis. I very much want to read about each of those examples, both good and bad – sounds like a good way to get an overview of the state of the art?
The document is 800 pages long (even if sparsely formatted), thus only for the truly dedicated.
I think it supports skimming and dipping into only the sections that sound interesting. If you’re comfortable with that, it’s worth a look even if you’re not super dedicated. And you can take “only” the first 250 pages as a standalone thing. The remaining 550 pages cover Ryan’s own works and progression in significant detail and you can pick and choose from that part.
But it’s a PhD thesis: it’s more of an argument for Ryan’s point of view than a general “overview of the state of the art”. For instance, it mentions Saga II and Mexica, but not the other works mentioned here as examples.
Thank you for the summary of Ryan’s work. I will take a deep dive in his thesis one day, but for now I’m just dipping my toes lest I drown in such monumental work.
I’ve always wondered about these so called drama managers. The way I see it, when building story generators one needs two modules:
1. Simulator. A world populated with agents who’s behavior changes the landscape.
2. Story sifter. A filter of the “debug log” generated by the simulator.
What would the function of a drama manager be? In general and in the context of the above mentioned architecture.