Concluding Remarks - Design and Analysis of Learning Classifier Systems - page 233

machine learning by using the latter to train LCS, and facilitates the good un-
derstanding of machine learning to improve the understanding of LCS. Overall,
approaching LCS from a different perspective has given us a clearer view of the
problems that need to be solved and which tools can be used to solve them.
This approach still leaves high degrees of freedom in how the LCS model itself
is formulated. The one provided in this work is inspired by XCS(F) and results
in a similar algorithm to update its parameters. One can think of a wide range
of other model types that can be considered as LCS but are quite different from
the one that was used here, one example being the linear LCS model that might
result in algorithms that are similar to ZCS. One thing, however, that is shared
by all of these models is what makes them an LCS: a global model that is formed
by a combination of replaceable localised models, namely the classifiers.
The model structure search itself might not have received the same attention
as common to LCS research. This was on one hand deliberate to emphasise
that, as defined here, finding the optimal classifier set is nothing else than an
optimisation problem that can be solved with any global optimiser. On the
other hand, however, it was only dealt with on the side due to the complexity of
the problem itself: most influential LCS research is contributed to the analysis
and improvement of the search for good sets of classifiers. Applying a genetic
algorithm to the optimisation problem results in a Pittsburgh-style LCS,as
in Chap. 8. Designing a Michigan-style LCS is a quite different problem that
cannot simply be handled by the application of an existing machine learning
algorithm. So far, such LCS never had a clearly defined optimal set of classifier
as the basis of their design. Such a definition is now available, and it remains
a challenge to further research how Michigan-style LCS can be designed on the
basis of this definition.
It needs to be emphasised that the model-centred approach taken in this work
is holistic in the sense that rather than handling each LCS component separately,
it allows us to deal with function approximation, reinforcement learning and
classifier replacement from the same starting point, which is the model.
Is taking this approach really so much better than the ad-hoc approach; that
is, does it result in better methods? This question can only be answered by eva-
luating the performance of a resulting LCS, and needs to be postpones until
such an LCS becomes available. Nonetheless, even the model-based perspec-
tive by itself provides a new view on LCS. Also, considering that most popular
machine learning methods started ad-hoc and were later improved by reformu-
lating them from a model-centred perspective, applying the same methodology
to reformulating LCS is very likely to be profitable in the long run.
Another question is whether theoretical advances in a field really help im-
prove its methods. It seems like that founding the theoretical understanding of
a method is a sign of its maturity. The method does not necessarily need to
be initially developed from the formal perspective, as Support Vector Machines
(SVMs) were [219]. Still, providing a theoretical foundation that explains what a
method is doing adds significantly to its understanding, if not also to its perfor-
mance. An example where the understanding was improved is the interpretation