model could look at a long sequence of rainy and sunny days, and analyze
the likelihood that one kind of weather gets followed by another kind.
Let's say it was found
that 25% of the time, a rainy day was followed by a sunny day, and 75%
of the time, rain was followed by more rain. Let's say
we found out additionally, that sunny days were followed 50% of the time
by rain, and 50% by sun. Given this analysis, we could generate a
new sequence of statistically similar weather by following these steps:
1) Start with today's weather.
2) Given today's weather, choose a random number to pick tomorrow's weather.
3) Make tomorrow's weather "today's weather" and go back to step 2.
What we'd get is a sequence of days like:
Sunny Sunny Rainy Rainy Rainy Rainy Sunny Rainy Rainy Sunny Sunny...
In other words, the "output chain" would reflect statistically the
transition probabilities derived from weather we observed.
This stream of events is called a Markov Chain. A Markov
Chain, while similar to the source in the small, is often nonsensical in the large.
(Which is why it's a lousy way to predict weather.)
That is, the overall shape of the generated material will bear little formal
resemblance to the overall shape of the source. But taken a few events at a time,
things feel familiar.
Doctor Nerve's Markov program uses English words instead of weather, yielding odd scrambled
results that sound like very disturbed English. The program looks at the likelihood
that every word you typed in gets followed by another word. After it generates all
these transition probabilities, it starts with the first word you typed and
starts generating random numbers to pick the sequence of words following it. You will
always get pairs of words that occurred in the original.
But looking at three or more words
at a time starts to look pretty bizarre.
The results are like dreamy crossed memories of the original text
whose phrases branch off and meander in simultaneously
unexpected and uncomfortably familiar ways.
For example, if you entered: "the boy and the dog went to the park."
The Markov chain might output: "the boy and the boy and the dog went to the boy"...note that the word "the" is the statistical pivot here - it was followed by "boy",
"dog", and "park", so there's a 33.3% chance that one of these words will follow "the"
every time the Markov Chain picks "the".
The Markov Chain would NEVER output: "boy the park."...because "boy" was never followed by "the" in the original story.
For more info about Markov Chains, refer to Scientific American June 89,
AK Dewdney's Computer Recreations column. Also, the book
"Computer Music - Synthesis, Composition, and Performance" by Dodge/Jerse
(Schirmer Books) is an excellent source of not only Markov analysis
techniques, but other statistical methods applicable to music as well.
Markov processes have been used to generate music as
early as the 1950's by Harry F. Olson at Bell Labs. Olson used them to
analyse the music of American composer Stephen Foster, and generate
scores based on the analyses of 11 of Foster's songs. Lejaren Hiller and
Robert Baker also worked with Markov processes to produce their "Computer
Cantata" in 1963. You can use Nerve's Markov program to do this by typing in
pitches and durations instead of words. For example, C#Q DQ C#Q DS EE. could
stand for the following melody: C# quarter note,
D quarter note, C# quarter note, D sixteenth note, E dotted eighth note. Use whatever notation
you want, just stay consistent - Markov will happily analyze your input
and generate statistically similar output.
-- Nick Didkovsky, Feb 16, 1996
Back to Nerve Interactive