Article about learning

[chuck at sanna.com]

Hello All,

While reading Gerry's latest paper, I was reminded of some things I've
been thinking about for a while.  They are all entangled with one another
so I'll just mention one and see what sorts out.

In ICMAUS models, there is an incoming data stream called New, and a data
structure called Old.  Old is built up from finding the best pattern
matches in New and Old (initially empty).

There can be many ways to look at New with one way representing the best
way (or at least the best way, given a particular method of search).  As
more New data comes in, I can imagine that a previously not optimal
compression may overtake the currently optimal one.

How does ICMAUS handle this?  How is Old modified to reflect the new
optimal compression?  Is it completely replaced by the new compression, or
does it evolve in a continous manner towards the better compression.

Is it a good idea to keep multiple compressions around, in case one of
them eventually wins out?  Maybe all these competing compressions are just
subsets of each other in some way and can be stored together in an 
efficient way.

Or maybe New should be completely re-examined whenever new data comes in.

Thanks for any thoughts on this.

Chuck