[r-t] A new Spliced Surprise Major canon

[Simon Gay]

I think this is a great idea, and evidently the results are already 
worthwhile.

A long time ago (probably around 1997) I remember asking, on the 
change-ringers list, whether anyone had tried using genetic algorithms 
for composition search. At that time it didn't seem as if anyone had 
looked into it, as far as I remember.

Your idea of using conventional search to recover from the introduction 
of falseness is very nice - and clearly you've shown that it works well 
in practice. It's a clever way of getting round the problem that truth 
is a global property, not something that you just want to optimise locally.

Simon



On 04/03/2013 08:04, Mark Davies wrote:
> I've been thinking for some time that more work needs to be done in 
> the field of "ordinary" spliced - by which I mean the straightforward 
> stuff with tenors together and familiar methods, the sort of thing 
> that gets rung every day. It seems most of the important contemporary 
> developments in spliced have focused on the (n-1) or (n)-part, 
> generally with cyclic part-ends. Of course we have seen some amazing 
> progress here, with Alan Reading providing us a fabulous example only 
> recently.
>
> However, sadly, many peal bands don't ring this sort of stuff, at 
> least not regularly. You're much more likely to see peals of 4, 5, 6 
> or 8-spliced in standard methods, and here the choice of composition 
> seems to lag far behind that of the single method. In fact, you could 
> say that not a lot has happened since Pitman produced his spliced 
> series in the 1940s and 50s: everything is based on CRUs, and it seems 
> difficult to source a composition with a good selection of little-bell 
> music.
>
> To my mind, that's unacceptable. OK, composing spliced is hard, and 
> Pitman was brilliant, but his compositions are sixty years old and 
> more, and, with the tools now at our disposal, we really ought to be 
> able to do better. So I've made a stab at setting a new standard for 
> the genre.
>
> My starting point was the idea that I could take a good, simple, 
> single-method composition, and add methods to it gradually, to produce 
> a series from "1-spliced" up. As the seed composition I chose my 5056 
> no.1 of Bristol Surprise Major. This has been superseded in absolute 
> music counts by later peals in my Bristol series, however I still have 
> a certain fondness for it, and it does have an elegance in its 
> structure which I thought deserved a new lease of life. It also has 
> very few calls, which gives a fresh new feel to the spliced arrangements.
>
> In order to "splice up" this basic calling I developed a set of tools 
> based on what Wikipedia terms "stochastic metaheuristics". This 
> appears to be a fancy name for a class of algorithms which use a 
> probabilistic approach to refine a solution through a series of steps. 
> Examples are the method of simulated annealing, stochastic tunneling, 
> and genetic searches. All are very different from the traditional 
> brute-force search, but have proved very successful when applied to 
> spliced.
>
> The key factor for success is choosing the correct scoring metric. For 
> example, in the genetic algorithm I breed compositions together by 
> randomly recombining nodes from two parents, and then, in successive 
> generations, filtering off all but the highest-scoring children. The 
> score must reflect the desire to maximise music and method balance 
> whilst keeping ATW counts, length and truth in check. The same is true 
> for other metaheuristic algorithms. It was quite remarkable how, in 
> building this new series, I would often find that tweaking the scoring 
> criteria was enough to open up new fields of previously-undiscovered 
> compositions. It was almost as if once I knew how to describe a 
> composition, I could find it.
>
> The other key breakthrough was finding a way to deal with the 
> discontinuities introduced by the axioms of changeringing. Falseness 
> is the primary problem. For any of these probabilistic algorithms to 
> succeed, it is necessary for them to be able to traverse valleys of 
> lower scores, in order to discover the high-scoring peaks. However, 
> the best peaks may be isolated by deep moats of falseness. If you 
> venture into these moats, by allowing the composition to run false, 
> you are unlikely get back out again. The search drowns.
>
> I addressed this major problem with what I termed a "deterministic 
> bridge". Continuing the analogy, as soon as the search's boots were 
> wetted by a step into falseness, I engaged a conventional brute-force 
> search to permute hitherto unaffected nodes in order to regain truth, 
> hopefully now on the other side of the moat. It proved to be enough to 
> implement this in one algorithm only - the method of simulated 
> annealing - and to engage it only when scores could not be maximised 
> further using true steps on "dry land".
>
> Further information about the series, and the figures for the 
> compositions, can be found here:
>
>     http://bronze-age.com/spliced/spliced1.html
>
> Note that the methods I have chosen are not primarily designed to 
> popularise a new "standard eight". (In fact, there are only six of 
> them for the time being!). I have tried to select examples which are 
> beautiful and musical in their own right, but they needed to be 
> familiar, too. For better or worse, I think many bands would be put 
> off by a collection of methods with completely unfamiliar names and 
> lines.
>
> So I stuck with Bristol and Superlative from the standard eight, and 
> added Cornwall and Lessness, both of which seem to be rung a lot these 
> days to single-method peals, and deservedly so. I also wanted a 
> quality example of a jx method, and here Deva fits the bill. It is 
> perhaps the least rung of all the methods I have chosen, but since it 
> has featured regularly on ringing-chat as AJB's favourite method, I 
> hope peal bands will at least have heard of it. With Bristol above and 
> a neat right-place below work, it's very pleasant to ring. Finally, 
> Malpas, from Chandler's 23, adds a touch of sophistication, perhaps.
>
> I hope this series will be of interest to the ringing community, and 
> that the time is right for something of its ilk. Comments, questions, 
> criticisms and suggestions welcome.
>
> MBD
>
> _______________________________________________
> ringing-theory mailing list
> ringing-theory at bellringers.net
> http://bellringers.net/mailman/listinfo/ringing-theory_bellringers.net