Statistical machine learning?

An article titled Serious Gaming in Georgia Tech University's Spring 2010 Horizons magazine says researchers are using games to improve tools for statistical machine learning. According to the article, researchers are "using gaming concepts such as narrative to help re-envision approaches to machine learning." This seemed interesting, so I Googled "game narrative" to see how it might differ from that of literature, for instance. This brought me to an except from Narrative, Interactivity, Play, and Games by Eric Zimmerman in which he says:

"I draw my definition from an essay by J. Hillis Miller: "Narrative," from the book Critical Terms for Literary Study (1995). Miller's definition of the term "narrative," grossly paraphrased, has three parts:

1. A narrative has an initial state, a change in that state, and insight brought about by that change. You might call this process the "events" of a narrative.

2. A narrative is not merely a series of events, but a personification of events though a medium such as language. This component of the definition references the representational aspect of narrative.

3. And last, this representation is constituted by patterning and repetition. This is true for every level of a narrative, whether it is the material form of the narrative itself or its conceptual thematics.

It's quite a general definition. Let's see what might be considered narrative according to these three criteria. A book is certainly a narrative by this definition, whether it is a straightforward linear novel or a choose-your-own-adventure interactive book, in which each page ends with a choice that can bring the reader to different sections of the book. Both kinds of books contain events which are represented through text and through the patterned experience of the book and its language.

A game of chess could also be considered a narrative by this scheme. How? Chess certainly has a beginning state (the setup of the game), changes to that state (the gameplay), and a resulting insight (the outcome of the game). It is a representation -- a stylized representation of war, complete with a cast of colorful characters. And the game takes place in highly patterned structures of time (turns), and space (the checkerboard grid).

Many other kinds of things fall into the wide net Miller casts as well -- some of them activities or objects we wouldn't normally think of as narrative. A marriage ceremony. A meal. A conversation. The cleverness of Miller's definition is that it is in fact so inclusive, while still rigorously defining exactly what a narrative is."

In any case, according to the Horizons article the researchers "see a strong connection between mathematical sequences and the process of narrative. Traditionally, machine learning focuses on immediate tasks rather than on sequences. It doesn't matter where you have been, it only matters where you are."

With something like G-code or robot programming in mind that sounds true.

The article next states, "But narrative can help make machine learning more flexible. With a story, it matters very much where you've been and those of us in machine learning have something to learn from that."

Therefore, there must be many similarities between machine learning theory and statistics?

Hide comments


  • Allowed HTML tags: <em> <strong> <blockquote> <br> <p>

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.