[r-t] Man and machine

[Don Morrison]

Mark Davies wrote:
 > I think a composer's brain eliminates nodes, as you say.

A fair bit of cognitive science research teaches us to distrust our 
introspective intuitions of what we're doing when solving mental problems, but 
I'm going to go out on a limb here, anyway.

I strongly disagree. I don't think that what I'm doing when attacking a 
compositional problem in my own head looks at all like what a traditional, depth 
first search a computer does, node pruning or not. It may be more closely 
related to some sorts of heuristic search, but possibly I only think that 'cause 
I understand the details of what really happens in heuristic search so little.

I'm pretty sure what I'm doing in my own head is much less focused and directed. 
I've typically got some chunks of composition I want use and find ways of 
joining up. A common line of thinking is "how can I get from this bit to any one 
of the remaining bits?" I'm shuffling things about willy-nilly, not in any 
disciplined order, going from this part of the composition to a completely 
different one and coming back again later. Thoughts for how to easily get from 
one place to another are often remembered from the distant past, possibly even 
with completely different methods.

After playing around with various bits and pieces I'll usually have an idea of 
what bits feel like I should be able to fit them together, and become frustrated 
with the difficulty of working out exactly how. The computer is much better at 
that, using exhaustive search to figure out quick ways of getting from one good 
place to another.

Curiously, the computer seems also to be particularly good at thinking "outside 
the box." I'll have various prejudices about what will or will not work for 
something, but because the computer may be "stupidly" pursuing all sorts of 
paths I'd never think worth following it may come up with something surprising; 
for example, a complex combination call that gets some music off the front with 
the tenors reversed or something. Having the computer blindly search for good 
music in a shorter length in a space so large it would be intractable at peal 
length sometimes produces such interesting new ideas, which can then be 
incorporated as required blocks at longer lengths using a more restricted search 
space.

It has long seemed a potentially worthwhile effort to figure out how something 
is done by hand, and try to automate that. But it never rises high enough up the 
priority queue, as more brute force searching is so much easier to implement, 
and continues to have more than enough interest and challenges to consume all 
the time available to work on it!

 > I think a computer search could be designed to make use of the
 > latter technique - it's a type of music pruning. It struck me after
 > discovering my "no duffers" Bristol Major composition that a
 > computer could have found this very quickly, if it adopted a pruning
 > method where branches of the search tree were abandoned if no music
 > had been produced after say 3 nodes. Generally this technique will
 > miss good compositions, but the compositions that it does produce
 > will have the benefit of a good distribution of music and no long
 > "duffer" stretches. This is definitely the next pruning optimisation
 > technique that I intend to program.

A similar scheme I've found quite useful is limiting the search to just good 
courses, plus only a little bit of potential linkage material. For example, in 
an F or Mx royal or max method limiting the search to only little bell courses 
between M and W, but anything at all between W and M. This isn't quite the same 
as having desirable blocks you're looking for ways to link up, since it can use 
or not use its choice of the good courses, however the search process takes it.


-- 
Don Morrison <dfm at ringing.org>
"Applying computer technology is simply finding the right wrench to
pound in the correct screw."                        -- author unknown