Archive: May 25th, 2008

Descriptive vs Generative models

[General,Research] (05.25.08, 9:31 pm)

Suppose we look closely at a system and build a model from it. That model will assuredly be useful in describing the system, and possibly predicting that system, but it is not necessarily the case that the model will allow us to build new variants or instances of the system. Examples of this are structuralist formulations of film or narratives. Valdimir Propp’s morphology of Russian folktales is a great example of this, the morphology may be used to analyze or describe folktales, but when reversed and used as a generational tool, far too much is missing. Other models may be used to generate artifacts. Examples of these can range from the work of Karl Sims, to CS paper generators, to any other abstract “generator” algorithm or program. These tend to be much more particular, applied to special cases and mini-domains rather than large scale categories.

There are other differences between generative and descriptive models. Descriptive models are generally much more useful for humans. Even if a descriptive model cannot be used to generate an artifact, it may still be used to check whether an artifact matches a given model. This is very helpful for human authors or creators, who might want to make something that mostly fits a model, but then perverts it, complicates it, or changes it in some interesting way. For instance, in the domain of superhero comics, Alan Moore’s Watchmen follows the overt tropes of the genre, but exposes a psychological depth that is otherwise not traditionally present. Other works might attempt to synthesize genre models, or for the sake of parody, will stress elements of the model to extremes. Descriptive models are enormously powerful as tools for creativity.

Generative models tend to be useful, not for human creators, but for electronic ones. A generative model has the structural elements of a system specified down to an algorithmic level, such that any algorithm processing machine may use the model to generate an artifact. Generative models must also be narrow enough for the algorithm to cohesively express a domain. With this level of precision, the model must take positions on how to express its domain. Value and content judgments are necessary, as decisions must be made regarding what to include or exclude, what to prioritize or emphasize. As a result generative models tend to much more resemble artifacts themselves, representing a certain creative take on a domain, rather than a broad model of the domain at large. The CS paper generatorhas a certain embedded model of what CS papers are like, and its approach serves to illustrate the absurdity of this model. Genetic Image has a similar understanding of how an image should be composed, and it is fundamentally and intrinsically limited by this viewpoint in terms of what it can express as a tool. As tools, constructive models are limited, but their value becomes apparent when they are seen as artifacts. Instead of critiquing a model through a carefully constructed parody, generative models critique or expose their own perspectives through their very execution.

From far enough away, both descriptive and generative models are similarly constrictive. The constraints of a structuralist descriptive model of a genre is as limiting as one that reduces the model to an algorithm. Furthermore, a descriptive model may be made into a generative model if it is focused enough, if one takes sufficient positions in their interpretation to resolve the inherent ambiguity of the descriptive model. Ambiguity is probably the key difference between the two. A descriptive model allows for a variety of interpretations, as such it is resistant to critique because of its open nature. A generative model has removed all ambiguity to the point where it may be executed algorithmically, but as a result may be critiqued and used as an expressive tool.