Databases Reference
In-Depth Information
the model performance is not good enough, further modification including
structure and parameter optimization would be required. Once the model
performance achieves the pre-defined criteria, the final model is produced.
4.3.2.
Incorporating knowledge into neurofuzzy models
As previously mentioned, neurofuzzy modeling has the advantage of combin-
ning expert knowledge with numerical data, helping not only to understand
the system but also to validate the model acquired from data. This section
presents a hybrid modeling method which incorporates knowledge-based
components, elicited from human expertise, into underlying data-driven
neurofuzzy network models.
33
In the modeling of engineering processes, there are two kinds of
information available. One is numerical information from measurements and
the other is linguistic information from human experts. The aforementioned
neuralfuzzy model is designed for data-driven models and cannot directly
deal with fuzzy information. To enable the model to utilize expert
knowledge presented by fuzzy if-then rules, an information processing
mechanism must be established.
The use of linguistic qualitative terms in the rules can be regarded as a
kind of information quantization. Generally, there are two different ways to
incorporate knowledge into neuralfuzzy models, as shown in Fig. 4.4. The
first one is to encode expert knowledge in the form of If-Then rules into
input-output fuzzy data, and then to use both numerical and fuzzy data to
train the neuralfuzzy model, as shown in Fig. 4.4(a). In cases where the data
obtained from the system are incomplete but some expert knowledge with
regard to the relationship between system input and output is available, this
method can be used to incorporate linguistic knowledge into data-driven
neuralfuzzy models. Fuzzy sets can be defined by a collection of
α
-cut sets
according to the resolution identity theorem. Linguistic information can be
represented by
α
-cut sets of fuzzy numbers. Expert knowledge represented
in the form of If-Then rules can be converted to fuzzy clusters in the
input and output spaces. The neuralfuzzy model can be trained using both
numerical data and fuzzy data which complement each other.
On the other hand, in many cases the knowledge that links the system
input and output is not available or not sucient to generate fuzzy relations
between system input and output. However, it is still possible to use expert
knowledge to improve model performance if some knowledge about model
