DM: TMC's Darwin

[Will Dwinnell]

"We're looking at a major investment in TMC's Darwin product.
Has anyone on the list had a chance to really put it to the test?"

To some degree, I have.  The last version I worked with would have 
been
whatever Thinking Machines was shipping in February (I worked at a 
beta
site), which would have been release 1.Beta 2_16 or possibly a little
later, so some of this may have changed.  At the time, Darwin 
consisted of
3 learning tools: a neural network, CARTŪ  and a k-nearest-neighbor 
system
(which TMC referred to as "memory based reasoning").  The neural 
network
struck me as fairly ordinary (backprop, conjugate gradient) and their 
CARTŪ 
system seemed to function adequately.  I felt that the k-nearest 
neighbor
(k-NN) system was not terribly useful in the big data mining 
environment
they are marketing to, and I told them so when I visited them in 
Bedford. 
k-NN, in its raw form, requires that all the training examples be 
carried
about for use during recall since the examples are the model.  This 
struck
me as completely unwieldy since data miners who would buy this kind of
software would have rather large databases of examples and 
implementation
of this sort of thing in code would get messy (at least, compare to
something like a neural network or a CARTŪ -derived decision tree).  I 
know
TMC had been making changes the entire time I used Darwin so maybe 
there
are more features now (such as clustering).

Will Dwinnell