The adjacent possible

This phrase was first coined, I think, by Stuart Kauffman, in his book “Investigations

The basic idea: Evolution may change speed, but it does not make big jumps. It typically progresses through numerous small moves, exploring adjacent spaces of what else might be possible. Some of those spaces lead to better fitness, some to less. This is low-cost exploration, big mutational jumps involve much more risk that the changes will be dysfunctional or even terminal in the immediate short term.[But for a counter view, read about “hopeful monsters“]

The same strategy may apply to many development programs, where big changes may require a very different set of human capital, whereas incremental changes would only require small changes in human capital requirements

Incremental searches for small improvements in a predictive model can be made in two ways:

  1. Testing the addition of new attributes, one at a time. This can be done in two ways:
    1. Manually, by clicking on one attribute at a time in the Design View and noting how it changes the performance of the current model. This could be described as a breadth-first search.
    2. Automatically, by clicking on “Most predictive of any kind” and then choosing “Find one additional attribute that gives the best performance”. This will enlarge the current model by one attribute. Repeating this process will expand the size of the current model by one attribute at a time. This could be described as a depth-first search
  2. Testing the effects of the removal of existing attributes of the design, one at a time. This is covered in detailed under Sensitivity Analysis

Caveat: If you are exploring the fitness landscape around an existing model, then adding an extra attribute can have two effects on performance. Firstly, say if you add an attribute that specifies a particular context, this is likely to reduce the coverage of the model (=TP/(TP+FN)). That is to be expected and not necessarily a problem. What matters is that within that more circumscribed context has the consistency of the model (=TP/(TP+FP)) increased or decreased?This additional attribute is in effect a scope condition.

For more on the idea of “the adjacent possible” see:

Spaces of the possible: universal Darwinism and the wall between technological and biological innovation. Andreas Wagner, William Rosen.  2014

Kauffman, Stuart A. Investigations. Oxford University Press,

Johnson, Steven. Where Good Ideas Come From: The Seven Patterns of Innovation.Where Good Ideas Come From: The Seven Patterns of Innovation. London: Penguin, 2011.

 “The Atlas of Economic Complexity.” MIT Press. Accessed January 12, 2016.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: