Corresponding errors etrain and egener are calculated
Page 495
brain, which the hypothesis of the goal-seeking neuron tries to explain (Freeman 1991; Kaneko 1989).
Adaptation is a feature of intelligence. Adaptation means ability of a system to change its structure and functionality according to dynamically changing environment for a better reaction. Adaptive intelligent information systems should be able to change their knowledge base during operation. For example, a fuzzy system should be able to change its rules, a neural network should learn new data "on the fly."
Two different ways to make space in a neural network in order to adapt to new data are: (1) to expand the structure if necessary by adding new nodes as new data come; and (2) to make the system forget about old data, thus making space for the new ones.
The following steps form a general methodology for evaluating the adaptation and forgetting in a connectionist model:
1. A data set is randomly divided into sections (sets) A, B, C, etc.
The following are steps for updating a fuzzy neural network structure FuNN (see figure 4.38 and section 5.3.4) through zeroing, in order to make the FuNN structure adaptable to new data and to control the adaptation/ forgetting phenomenon: (1) A FuNN architecture is initialized (either randomly or according to an existing set of initial rules). (2) The FuNN is trained with data. (3) Fuzzy rules are extracted from the trained FuNN. for example by using REFuNN algorithm (a threshold is set in advance, e.g., 0.7). (4) The FuNN architecture is then updated according to the extracted fuzzy rules by zeroing the connection weights which are less than the threshold (or their absolute values are less than the threshold if NOTs are used in the extracted fuzzy rules). A new updated FuNN is