> I know this has been discussed ad nauseum, but I would like to know a good
> central place where the Michigan Approach and the Pitt (or Pittsburgh)
> approach are compared. Any I've read write on the basic workings and
> comparison, but I think these are lacking in depth, because they are often
> short accounts of the differences.
> I must be missing out on years of debate (between proponents of both
> approaches) because what I've found so far seems to miss the depth of the
> arguments for and against both approaches and their relative merits.
I don't know of any centralised place where the Pitt vs Michi approach
is discussed. You might try asking people on comp.ai.genetic
I had some lecturers on LCS by a researcher (Larry Bull) in them and
he was interested in creating new types of LCS or exploring extensions
to simpler LCS systems such as Zeroth Classifier Systems. Which are
based on stripped down michigan style classifier systems.
Another of the things he was interested in was a neural classifier
system. Where the representation was interpreted as a recurrent Neural
net, but searched by a modified genetic algorithm. This can be seen as
a form of Pittsburgh Classifier system as the neural net could
represent many rules and have memory in itself.
So from what I have seen of the LCS community would say the debate has
died down and researchers accepted both have advantages and
disadvantages and use whatever works dependant upon the type of system
they are trying to create.
My view on LCS is that it is too restricted in how rules are created
to have much future in general problem solving (my own research I am
slowly trying to cross Pushpop (a form of GP), LCS and Tierra).