[r-t] A new Spliced Surprise Major canon

Mark Davies mark at snowtiger.net
Mon Mar 4 21:14:23 UTC 2013

Don Morrison asks,

> Another global property that I would think might introduce similar
> problems to falseness is keeping things ATW. Did you have to do
> anything special for it, or does just maximizing the number of
> bell/place bell pairs as part of your overall scoring just usually
> lead to ATW compositions?

You have hit the nail on the head, Don. Initially I had a simple ATW 
flag, which added a constant to the composition score; but this 
introduces a hard discontinuity, making it unlikely the search will move 
towards ATW or, if it does discover ATW, escape from the first such 
local minimum with that property.

So I switched to using a count of the number of missing bell/place bell 
pairs, just as you supposed. Smoothing the scoring out like this worked 
much better. I did something similar for method balance: the scores have 
to gradually improve as you get better balance. (I wonder now if 
something similar could be done with "degrees of falseness".)

I also found I had to vary these criteria for the specific composition I 
was working on, and the specific stage I had got to with it. For 
example, for some peals it was necessary to search first for 
compositions with good method balance, and only when no further progress 
could be made, tune the scoring criteria to prioritise better music. 
After maximising music one could sometimes return to method balance and 
make new progress there, whilst maintaining music counts.

Also at a particular stage one algorithm would be more fruitful than 
another, so I'd need to swap back and forward between them. It was quite 
instructive to find out just how much things could be improved by 
chipping away like this, getting stuck, making progress with a approach, 
getting stuck again but then finding an earlier setup could now forge 
ahead. In this way the task became one of collaborative exploration, 
with man and machine working together to refine the results ever further.

So these peals are definitely not the result of simply plugging in a 
piece of data at one end and watching the results spool out the other. 
In fact the whole process was much more fun than that!


More information about the ringing-theory mailing list