|
The tuning stage is usually one of the most complex tasks when designing fuzzy
systems. The system behavior depends on the logic structure of its rule base
and the membership functions of its linguistic variables. The tuning process is
very often focused on adjusting the different membership function parameters
that appear in the system definition. Since the number of parameters to
simultaneously modify is high, a manual tuning is clearly cumbersome and
automatic techniques are required. The two learning mechanisms most widely used
are supervised and reinforcement learning. In supervised learning techniques
the desired system behavior is given by a set of training (and test)
input/output data while in reinforcement learning what is known is not the
exact output data but the effect that the system has to produce on its
environment, thus making necessary the monitoring of its on-line behavior.
The Xfuzzy 3 environment includes four tools for this design stage:
xfdm and xftsp
are knowledge acquisition tools. The first one allows obtaining the structure of
inference systems used as fuzzy approximators or classifiers, while the second one
is specially focused on time series prediction applications.
xfsl is a parameter adjustment tool based on the use
of supervised learning algorithms. In supervised learning techniques, the desired
behavior of the system is described by a set of training (and test) patterns.
Supervised learning attempts to minimize an error function that evaluates the
difference between the actual system behavior and its desired behavior defined by
the set of input/output patterns.
Finally, xfsp is a simplification tool that allows
reducing the number of membership functions and compacting the rules bases of a
fuzzy system to facilitate its software or hardware implementation and to increase
its linguistic interpretability.
|
