Extracting Rules from Data

In theory, neural networks and fuzzy systems are identical in that they are convertible, but in practice, each has its own advantages and disadvantages. In the case of neural networks, knowledge is not automatically invented by the back-propagation algorithm, but the learning process is relatively slow and the texture of the trained network is difficult. Meanwhile, since the input space of the fuzzy system must be divided into fuzzy regions, it is very difficult to wield the fuzzy system to a problem with a large number of input variables. Moreover, difficulties arise in conquering knowledge through expert interviews. However, in the case of fuzzy systems, once knowledge is gained, how the system works is relatively understood. Several methods have been developed to extract fuzzy rules from numerical data to fill the knowledge conquest gap between the two technologies. One Tideway uses a neural network to iterate over fuzzy rules. For example, neural networks are extended to automatically periscope fuzzy rules on numeric data. The main limitation of this method is that you have to specify the number of divisions for each input variable in advance. It also requires splitting of the output variable, but this is not a downside because it can be an unshakable hand depending on the property accuracy. Flipside Tideway extracts fuzzy rules directly from numerical data. For example, fuzzy rules are derived by dividing the input space into fuzzy regions and the output space into regions, and by determining the fuzzy region containing each numerical input data and the output region containing each output data. Fuzzy rules with variable fuzzy regions are extracted for naming problems. This tidal flat solves the problem of the purge system described above. The entry area for each admission is displayed as a series of hyperboxes; overlap of hyperboxes between each other for the same admission is allowed, but overlap between different classes is not. If the data is in the hyperbox belonging to the class, it is judged as that class. The learning algorithm dynamically expands, splits, and contracts hyperboxes when multiple classes overlap.