Does the MM really have to be reserved for prediction? Supposing a
series of flourishes, agents, programs, what have you, assembled as a
graph; and some number of these traversed manually, or otherwise, to
approximate a gesture. Perhaps one simply trains the MM to attempt
subsequent traversals, adding new or sufficiently similar material?
Then the short-range leanings of MMs are preserved, positioning the
fitness function as assessor and perhaps mutator of the MM, at
whatever appropriate point in the foodchain.
Granted, in this scenario you'd be abstracting to a pretty high level
the decision-making of the MM, which in turn might require a
sympathetically broad model of musical gesture.
Thanks for comment,
Al
On Wed, Oct 12, 2011 at 2:56 AM, Dan S <danstowell+toplap_at_gmail.com> wrote:
> 2011/10/12 David Griffiths <dave_at_pawfal.org>:
>> On Tue, 2011-10-11 at 08:23 +0100, Dan S wrote:
>>> For musical interestingness, the idea of the Wundt curve works pretty
>>> well - you need to be unpredictable but not too unpredictable since
>>> unpredictable is just noise. So one thing you could do for discrete
>>> melodies, if they're not too short you could use something like a
>>> markov model to predict the later parts of the melody based on the
>>> early part. If the trained markov model improves prediction relative
>>> to a uniformly-distributed markov model, but not to perfect
>>> prediction, maybe that's useful...
>>
>> So you'd train the markov model on the style of melody you're interested
>> in, and the success of a prediction of part of the output is a metric
>> for the program that generated it? Sounds good.
>
> That's not what I had in mind but it's another option! Markov models
> are quite limited (no overarching structure) so if you trained them on
> a particular known style, it'd learn short-range tendencies but not
> really longer-range. I think the MM would be more useful for driving
> evolution by analysing its results: e.g. for a given population you've
> produced, train a markov model on the whole lot, then apply the idea
> of the wundt curve - the winning individuals are the ones which the
> markov model rates as unpredictable, but not completely unpredictable.
> No idea if it'd end up with good results, but that's the magic of
> GAs...
>
> Dan
>
>> It would also be possible to generate markov models of betablocker
>> programs themselves, as all possible programs are executable - but
>> that's something else again :)
>>
>> cheers,
>>
>> dave
>>
>
--
Al Matthews
Received on Wed Oct 12 2011 - 13:25:43 BST