# Curating Simulated Storyworlds — Chapter 13: Conclusion

> NOTE: THIS IS A MARKDOWN CONVERSION INTENDED FOR LLM CONSUMPTION.
> IF YOU ARE A HUMAN BEING, READ THE PDF INSTEAD:
> https://jamesryan.computer/papers/2018_ryan_curating_simulated_storyworlds.pdf
>
> The PDF is the canonical document, typeset with every figure and table. This file
> is a plain-text conversion of the LaTeX source, published HERE so that software agents
> can read the thesis without parsing a PDF. It is lossy and is not meant for humans.
>
> This file is chapter 13 of 14 only: Conclusion (~2,700 words).
> See below for links to markdown conversions of the individual chapters.
>
> **Structure.** Sections are `##`/`###`/`####` headings. Cross-references are resolved to
> their numbers ("Section 3.1.1", "Chapter 5"), so a heading can be searched for directly.
> Citations are inline bracketed keys such as `[ryan2017grimes]`; every key used in the text
> has an entry in the bibliography, formatted as `- **[key]** Author. Title. Venue, Year.`
> Footnotes appear inline as `[Note: ...]`.
>
> **Conversion limits.** Figures and most complex tables appear as `[Figure omitted]` or
> `[Table omitted]`. Code fences carry no language tag. A few diacritics in author names
> were simplified from the Latin-1 source.
>
> **All files.**
>
> Index (this list, plus the abstract; small): https://jamesryan.computer/thesis/index.md
> Whole thesis in one file (large, ~1.6 MB): https://jamesryan.computer/thesis/full.md
>
> By chapter:
> - 01. Introduction (~4,600 words, ~30 KB) — https://jamesryan.computer/thesis/01-introduction.md
> - 02. Defining Emergent Narrative (~6,900 words, ~47 KB) — https://jamesryan.computer/thesis/02-defining-emergent-narrative.md
> - 03. The Pleasure of Emergent Narrative (~21,000 words, ~135 KB) — https://jamesryan.computer/thesis/03-the-pleasure-of-emergent-narrative.md
> - 04. The Pain of Emergent Narrative (~41,000 words, ~268 KB) — https://jamesryan.computer/thesis/04-the-pain-of-emergent-narrative.md
> - 05. Refining Emergent Narrative (~11,000 words, ~73 KB) — https://jamesryan.computer/thesis/05-refining-emergent-narrative.md
> - 06. James Ryan Generator (~2,800 words, ~16 KB) — https://jamesryan.computer/thesis/06-james-ryan-generator.md
> - 07. World (~20,000 words, ~129 KB) — https://jamesryan.computer/thesis/07-world.md
> - 08. Case Study: Diol/Diel/Dial (~9,300 words, ~59 KB) — https://jamesryan.computer/thesis/08-case-study-diol-diel-dial.md
> - 09. Talk of the Town (~25,000 words, ~162 KB) — https://jamesryan.computer/thesis/09-talk-of-the-town.md
> - 10. Case Study: Bad News (~22,000 words, ~132 KB) — https://jamesryan.computer/thesis/10-case-study-bad-news.md
> - 11. Hennepin (~32,000 words, ~209 KB) — https://jamesryan.computer/thesis/11-hennepin.md
> - 12. Looking Ahead: Sheldon County (~11,000 words, ~75 KB) — https://jamesryan.computer/thesis/12-looking-ahead-sheldon-county.md
> - 13. Conclusion (~2,700 words, ~17 KB) — https://jamesryan.computer/thesis/13-conclusion.md
> - 14. Bibliography (~29,000 words, ~233 KB) — https://jamesryan.computer/thesis/14-bibliography.md

I would like to conclude this earliest dissertation in computational media by first outlining the contributions that I have made in its course, with the hope that I have done well through these contributions to express the myriad concerns that are central in this nascent academic discipline. To wit, in this document I have made the following contributions:

-  *A definition of emergent narrative.* After some deliberation, I converged on the following definition of ‘emergent narrative' in Chapter 2: ‘narrative that emerges out of computer simulation of character activity, or the methodology of generating narrative in that way'. My aim here was to join three distinct traditions that I view as instantiating the same form: the simulationist school of noninteractive story generation, as exemplified by systems like *Saga II*, Sheldon Klein's murder-mystery generator, and *Tale-Spin*; videogames driven by emergent character simulation, such as *Dwarf Fortress* and *The Sims*; and the bottom-up approach to academic interactive storytelling that is exemplified by the ‘emergent narrative' research programs of Ruth Aylett, Sandy Louchart, Mariet Theune, Ivo Swartjes, and their respective collaborators. As such, by my ecumenical definition, emergent narrative may obtain in both interactive and noninteractive works, as long as uninhibited character simulation is at work.

-  *An apologetics for emergent narrative.* While others have argued extensively about the merits of interactive emergent narrative, in Chapter 3 I worked to unpack the essence of character simulation itself. In doing so, I identified analogies of the form to nonfiction, lived experience, worldbuilding, and art brut. In particular, I agued at length that emergent narrative actually works more like nonfiction than fiction; here, my approach was to review a series of definitions of nonfiction, which I then applied to a work of curationist emergent narrative—Tim Dinee's *Oilfurnace* [denee2010oilfurnace]—to show that the definitions hold for that work. Finally, I ended the chapter by identifying what I call *aesthetics of emergent narrative*.

-  *A polemic against emergent narrative.* In addition to providing what is perhaps the most extensive defense of emergent narrative, I also contributed what may be the harshest critique of the form to date. In Chapter 4, I extended criticisms of emergent narrative advanced by various writers including Noah Wardrip-Fruin and Espen Aarseth to identify a series of pitfalls at the levels of both simulation and curation. Most primarily, I argued that simulation crafters often fail to enable emergent story structure, and that such structure can only truly manifest through a procedure of curation.

-  *A refined approach to emergent narrative.* This contribution came in the form of my framework for curationist emergent narrative, which was the subject of Chapter 5 and of this dissertation more broadly. Here, I proceeded from the material contained in both my apologetics and polemic for the form—my challenge to myself was to devise a refined approach to emergent narrative that takes its harshest appraisals as an intellectual basis.

-  *A set of artifacts, in the form of three simulation engines and three media experiences.* These contributions were constituted in my reporting of the completed simulation engines *World* and *Talk of the Town* in Chapters 7 and Chapter 9, the completed media experiences *Diol/Diel/Dial* and *Bad News* in Chapters 8 and Chapter 10, and the in-process engine *Hennepin* and the in-process experience *Sheldon County* in Chapters 11 and Chapter 12.

-  *A set of case studies in curationist emergent narrative.* For each of the three reported media experiences (and in the corresponding chapters), I provided case studies that discussed them in terms of the curationist framework at the heart of this dissertation. In delivering these case studies, my goal was to provide an actionable account both of how curationism may be carried out and why that is worth doing. Toward this aim, I specifically identified a series of design patterns at the levels of both simulation and curation, including: contingent unlocking, causal bookkeeping, sifting patterns, and sifting heuristics. This contribution answers Ian Horswill's call for the utilization of the *case* as a reporting mechanism in computational media [horswill2013science].

-  *An autoethnography of my simulation and media practice.* Jonathan Lessard, another member of my reading committee, encouraged me to provide in this thesis an *autoethnography* recounting the evolution of my practice, and indeed this became one of its contributions. Across the openings of the chapters in Part II, I told the story of how my simulation and media practice, and the curationist framework, all coevolved over a six-year exploration of computation and research.

-  *A history of emergent narrative.* While this dissertation was not an historical study, I did provide a significant historical account in the form of brief notes and mentions distributed across the document, especially in its footnotes. This included my tracing of the origins of the term ‘emergent narrative' in Chapter 2; the first substantial account of the earliest known work of emergent narrative, *Saga II* (1960), in Chapter 4; an excavation of Sheldon Klein's murder-mystery generator, also in Chapter 4, which worked to show how the system has for decades been misunderstood (this was in the style of Noah Wardrip-Fruin's study of *Tale-Spin*); and numerous other details.

I would now like to articulate some brief thoughts on a notion that is bigger than curationist emergent narrative, but to which it still pertains: *generators that understand their outputs*. Earlier in this dissertation, I referenced researcher Mike Cook's pronouncement that “the aesthetics of big numbers is dead”. In the original context of that death knell—it appeared in a *New Scientist* article about boredom with the videogame *No Man's Sky* [games2016no]—writer Douglas Heaven also quotes researcher Gillian Smith:

> A vast computer-generated playground may instill awe, but size alone is not enough to sustain interest. “What actually gets us to engage with something is how well-crafted it is and how well it resonates,” says Smith. [heaven2016when, n.p.].

 Around that same time, Kate Compton wrote an influential blog post identifying what she calls the *10,000 Bowls of Oatmeal problem*:

> So your algorithm may generate 18,446,744,073,709,551,616 planets. They may each be subtly different, but as they player is exploring them rapidly, will they be *perceived as different*? I like to call this problem the *10,000 Bowls of Oatmeal* problem. I can easily generate 10,000 bowls of plain oatmeal, with each oat being in a different position and different orientation, and *mathematically speaking* they will all be completely unique. But the user will likely just see *a lot of oatmeal*. *Perceptual uniqueness* is the real metric [compton2016so, p. 9]

 In a jam-packed short paper on the topic, Rogelio Cardona-Rivera terms this the *Kaleidoscope Effect* and calls for the development of a poetics for *cognitively-grounded* procedural content generation [cardona2017cognitively]. Relatedly, Emily Short has advocated for generated content that is *meaningfully* correlated to a storyworld or to gameplay [short2016bowls]. Expounding on the issue himself, Mike Cook called for a more considered language of generative design [cook2016alien], which Isaac Karth has now supplied in a brilliant paper on the *poetics of procedural generation* [karth2018preliminary].

More broadly, in procedural generation, this decade has been marked by a growing emphasis on building generators that are more understandable to their designers [craveirinha2013author,cook2016danesh,zook2016automated,shaker2016evaluating]—here, Gillian Smith and Jim Whitehead's 2010 paper “Analyzing the Expressive Range of a Level Generator” is the seminal work [smith2010analyzing]. [Note: In a clever inversion, Antonios Liapis and collaborators aim to make *designers more understandable to their generators* [liapis2014designer].] Now, a new prospect is emerging: generators that understand themselves, or more specifically, their own outputs [smith2017value]. This is the subject of Part III of Adam Summerville's dissertation [summerville2018learning], which provides an array of techniques that may allow generators to more effectively reason about their own outputs.

In this document, I have introduced a framework, curationist emergent narrative, that suggests a related phenomenon: simulated storyworlds that understand themselves, so that they may tell their own stories. I have been calling for exploration of this notion since 2015, when in a conference paper I asked, “How does one make a system that can discern stories embedded in the morass of data that its simulation produces?” [ryan2015open, pp. 20–21]. That paper would ultimately become the seed for this dissertation, and with this document I believe that I have now answered the question to some extent. In articulating my curationist framework, and particularly its architecture, my aim has been to provide an actionable account of how to build systems that are able to tell stories about what happens in simulated worlds. As this possibility becomes viable, an array of hitherto impossible media experiences are enabled, and the cutting edge of computational media is pushed forward.

But how viable *is* this possibility, at least right now? Let us turn to an excerpt from Jonathan Lessard's comments on an earlier draft of this thesis:

> To retrieve interesting stories, you will likely need to author story patterns or templates for the system to parse the chronicle: “a long lost treasure found”, “revenge for humiliation”, “the love triangle”, etc. And the system will find them. And as is usually the case with templates, readers will soon see through them and go: “Oh, here goes the revenge for humiliation story, except with Bob instead of John”... How *really* different is that from a system that begins with these templates and then generates a story to fit it, the exact approach you contest throughout your dissertation? You will say: “the difference is that it *really* happened”... OK, but if I let a generator go until it gives me exactly what I wanted in the first place, how *real* is that? [Note: Personal communication, July 27, 2018. Adam Smith made a similar point at my thesis defense.]

The general problem here is that a simulated storyworld cannot tell its own stories without harnessing some proceduralized method for doing so. As Jonathan indicated in his critique, one can define *sifting patterns* for retrieving certain kinds of material, but then the computer can only excavate material that matches those particular patterns. This is the impetus for *sifting heuristics*, as I explained earlier on, which is a notion that Jonathan connects to the historiographic practice:

> Historians don't approach their material looking for a pattern or a template. They often think in terms of “letting the archive speak”—that is keeping a mind open to serendipity. Of course, their *subjectivity* and *context* will inform what they find, but that is also what the “sifting patterns” miss. Since context and subjectivity change, different historians at different times from different places find different stories *in the same archives*! [Note: Ibid.]

 Thus, for a computer to sift out material that enables not only a considerable variety of stories, but moreover ones that may surprise humans and faithfully express the *expressive range* [smith2010analyzing] of an underlying simulation, we will likely need to develop robust sifting heuristics. This is something that is not delivered in this thesis, because I do not personally know how to do this yet, and it is a very hard challenge. I would like to issue a call to others to roll up their sleeves and get to work on this critical problem in emergent narrative—indeed, sifting heuristics could likely be the topic of another entire dissertation.

Let us look then to a possible future, one in which an array of sifting heuristics have been developed and implemented in actual computational systems. In this future, would simulated worlds finally be able to tell their own stories?

Sadly, we encounter here a more nefarious problem, one that is potentially insurmountable—let us call it the *narrativization bottleneck*. Even if a system can employ sifting heuristics to excavate material that does not fit any defined pattern, but rather matches an abstract (and potentially novel) idea for a story, that is only half the battle. Once the material is extracted, there is still work to be done in order to produce an actual story—this is the task that I call *narrativization*. And it is here that we face a critical bottleneck: even if an essentially infinite diversity of stories can emerge in a simulated storyworld, and even if the system can employ sifting heuristics to extract a corresponding diversity of material, the computer will not be equipped with the means with which to construct actual stories out of all that material. Save for an automated system that can invent its own mechanisms for doing so, the existence of such methods will be predicated on human labor. But then we are back to the same fundamental problem: a world can only tell its own stories if a human has told it how. This is the narrativization bottleneck, and it is particularly devastating here because it works to constrain emergence, and the aesthetics of emergent narrative are diminished in turn.

How can we overcome the narrativization bottleneck? It would require something like the equivalent of sifting heuristics, but for the task of building actual stories out of raw material. While it is quite reasonable to imagine such an automated procedure in the domain of story sifting, it may not be within reason here. To be clear, this would require a computational procedure for telling *arbitrary* stories. While we find in the half century of work on *story generation* an array of methods for getting computers to tell stories, there is no system that can tell *arbitrary* stories. Story generators tell the stories that their human creators have built them to tell. To tell an arbitrary story is something that only humans can do, and in fact, doing so as a machine would likely require *artificial general intelligence* [goertzel2007artificial]. If there is a more tractable solution, I would love to hear about it, but this is a problem that could take a thousand dissertations to solve.

Sheldon Klein once wrote in a conference paper, “It is a rather long distance from the promise of the title to the fragmentary output just presented” [klein1971program, p. 30]. While this thesis makes a number of contributions—I have already enumerated them—I am not able to provide a solution to the narrativization bottleneck, and it is possible that no such solution will ever come to exist.

If that is the case, then what are we left with? Let me again quote from Jonathan Lessard's feedback on an earlier draft of this thesis:

> My own conclusion to your work is that the automatic curation and narration is likely to be an impasse (though I'm always open to be surprised) whereas we *know* that human curation can be delicious. We know this from Dwarf Fortress, we know this from Bad News, we know this from The Sims. Reading about your simulations was a real delight. I wanted to find amazing stories for myself like the ones you have chosen to narrate to us. And indeed, once I became proficient enough with Talk of the Town, I had a great time finding interesting stories for myself (who is this bartender who used to be a doctor after being a fireman???) However, these simulations are difficult to parse for oneself. But what if you provided us with curation affordances? Then *we* would find the stories that interest *us* and we would surprise *you* with what we find.

 Regardless of the ultimate prospects for automatic curation, human-powered curation is not only feasible, but in fact it is already thoroughly demonstrated. As I argued at length in Chapter 5, it is how successful works of emergent narrative have worked all along. So how can we make it work even better? Incidentally, I think I have already answered that question.

All of the curationist steps leading up to narrativization do not have to be utilized as part of a fully automated pipeline. Alternatively, the corresponding modules could be framed as the components of a *curation support framework*: a computational system that is intended not to obviate human curation, but rather to *facilitate* it. In this way, even if arbitrary narrativization is impossible, we can still build full-fledged computational curationist systems, but ones in which the narrativization step is offloaded to human storytellers. In fact, this might actually be the most viable architectural pattern for curationist emergent narrative—perhaps you ought to try and find out.

{*}