Using the Neurofuzzy Approach to Building Dependable Neural Network Systems
Brown, M., Abed, Z. and Harris, C.J. (1995) Using the Neurofuzzy Approach to Building Dependable Neural Network Systems. Neural Networks - Producing Dependable Systems , 4.2.1--4.2.9.
Full text not available from this repository.
Novel computing techniques such as neural networks and fuzzy logic have been slow to make a significant impact on British industry, despite the fact that the Japanese and Far Eastern countries have been eager to embrace both technologies. Fuzzy and neural systems can be found in many household electrical appliances and the word ``fuzzy'' has come to signify quality, reliability and performance to the Japanese people. The Laboratory for International Fuzzy Engineering research (LIFE) has just completed its 6 year programme in Japan, and this venture was supported by about 50 of the top electrical and automotive companies which ensured a rapid transfer of technology from the University research laboratories to industry. The approach taken by these researchers is that thorough testing is a valid way to assess and certify their systems and this is exemplified by the fuzzy-controlled Sendai subway which was developed in 1985 and then tested for approximately 2 years before being certified. The nonlinear fuzzy control techniques used by Hitachi, ensured that the passenger ride comfort was improved, the automatic stopping system was more accurate and overall and it used less fuel. Researchers in the West have been concerned with the theoretical analysis of the stability of such systems and while this has produced some interesting results in recent years, the final certification of these systems is always done by thorough testing (simulation and real-world). Therefore, for engineers to adopt these techniques, the design cycle must be as transparent (the ability to understand the relationships contained in the network's structure) as possible, allowing both the integration of various knowledge sources (data, rules, polynomial-type relationships) into the overall system and the use of ``standard'' data and network analysis tools. Contrary to popular opinion, many neural and fuzzy systems have a fairly simple structure and therefore can be analysed and insights gained into the knowledge stored in the ``weights''. However, the different systems offer varying levels of transparency and this paper will be concerned with the so-called neurofuzzy systems that combine the learning abilities and structure of a neural network with the rule-based representational abilities of a fuzzy system. This paper is concerned with illustrating how various type of knowledge can be integrated into a neurofuzzy system and with showing how constructive learning algorithms can be used to automatically extract this information directly from a data set. Therefore, the fuzzy rule-based representation can be used for prior knowledge initialisation or for posterior analysis using a natural language interface. This can potentially increase the dependability of a system as it allows the designer to analyse the structure of the trained network and correct any deficiencies that may have existed in the training data. Rather than viewing the product development as a one-shot operation, these techniques support an iterative design-cycle methodology, where the system's overall performance is gradually improved and the natural language interface recognises that a human developer/expert is central to the product's design.
|Item Type:||Conference or Workshop Item (UNSPECIFIED)|
|Additional Information:||Organisation: ERA|
|Divisions:||Faculty of Physical Sciences and Engineering > Electronics and Computer Science > Comms, Signal Processing & Control
|Date Deposited:||04 May 1999|
|Last Modified:||27 Mar 2014 19:51|
|Further Information:||Google Scholar|
|RDF:||RDF+N-Triples, RDF+N3, RDF+XML, Browse.|
Actions (login required)