Information Technology Reference
In-Depth Information
where the output is of interval scale. That also changed the perspective of how
LCS handle sequential decision tasks: they act as function approximators for
the value function that map the states and actions into the long-run reward that
can be expected to be received when performing the action in this state, where
the value function estimate is updated by reinforcement learning. By replacing
classifiers in the population, LCS aim at finding the best representation of this
value function [138].
1.2.2
Applications and Current Issues
Learning Classifier Systems are applied in many areas, such as autonomous robo-
tics (for example, [75, 100]), multi-agent systems (for example, [87, 61]), econo-
mics (for example, [221, 169, 3]), and even trac light control [39]. Particularly
in classification tasks, which are supervised learning tasks where the output is
of nominal scale, their performance has been found to be competitive with other
state-of-the-art machine learning algorithms [98, 152, 7].
Nonetheless, even modern LCS are not free of problems, the most significant
being the following:
Even though initially designed for such tasks, LCS are still not particularly
successful in handling sequential decision tasks [11, 12]. This is unfortunate,
as “there is a lot of commonality in perspective between the RL commu-
nity and the LCS community” and more communication between the two
communities would be welcome [149].
Most LCS feature a high number of system parameters, and while the effect
of some of them is ill-understood, setting others requires a specialised know-
ledge of the system. XCS, for example, has 20 partially interacting system
parameters [57].
No LCS features any formal performance guarantees, and even if such gua-
rantees might not always seem particularly important in applications, the
choice between a method with such guarantees and an equally powerful me-
thod without them will be for the one that features such guarantees.
There is no knowledge about the assumptions made about the data, and as a
result there is also hardly any knowledge about when some LCS might fail.
Very few direct links between LCS and other machine learning methods are
established, which makes the transfer of knowledge for mutual gain hard, if
not impossible.
The general lack of rigour in the design of LCS leads to a lack of their
acceptance in the field of machine learning. Together with the previous point
this inhibits the exchange of ideas between possibly closely related methods.
These problems concern both practitioners and theoreticians, and solving them
should be a top priority in LCS research. Many of them are caused by designing
LCS by an ad-hoc approach, with all the disadvantages that we have descri-
bed before. This was justified when insucient links were drawn between LCS
and other approaches, and in particular when the formalisms were insuciently
 
Search WWH ::




Custom Search