|
|
|
|
From a practical point of view, all models of expertise have their place and possess special features designed to represent specific types of knowledge. Knowledge engineering research has produced a wide range of models of expertise, which have been classified into five major classes:
Heuristic models: offer a symbolic, or representational, description of expertise that is seen as a combination of facts and heuristics. A rule-based or heuristic system would be an excellent choice, for example, when the hierarchy between facts and decisions is well established;
Deep models: suggest that experts use a variety of deep knowledge structures represented differently from heuristic rules. Deep models of expertise are said to better represent the expert functions at each level of the expertise model hierarchy;
Implicit models: offer a representational description with a particular emphasis on the implicitness or explicitness of knowledge. Implicit models have become increasingly popular with the development of artificial neural networks;
Competence models: these models are representation independent and principally distinguish between domain and task knowledge. The focus of competence models is usually on the strategy of a problem-solving process. These models provide an extension of a functional rule-based system from which the heuristic knowledge was extracted to generate an explicit knowledge level description of a specialist
Distributed models: distributed systems can be compared to negotiation schemes. The tools implementing this approach are still on the research bench, as many aspects of fuzzy systems development for which fuzzification and defuzzification remain both attractive and mysterious to apply.
While some of these programs were relatively modest, others were quite ambitious and important both in scope and in funding. The advantages and limitations of using ES technology were analyzed in great detail in one of the first reported efforts on combating corrosion with ES.The Stress Corrosion Cracking (SCC) ES (SCCES) had been created to calculate the risk of various factors involved in SCC, such as crack initiation, when the user supplied evidence.The main goal of this effort was to support the decision process of general materials engineers.
A modeling neural network essentially acts as a mapping operator or a transfer function, taking inputs normally fed to a process and computing its predicted output values. Since the input output mapping can be either static or dynamic, the technology can be applied to a broad range of applications, being particularly efficient with real-time operations. But neural networks are also well adapted to perform other implicit expert functions such as pattern recognition and classification. An artificial neural network (ANN) is a network of many very simple processors or neurons (Figure 22), each possibility having a small amount of local memory.
The interaction of the neurons in the network is roughly based on the principles of neural science. Unidirectional channels that carry numeric data based on the weights of connections connect these neurons that operate only on their local data and on the inputs they receive via the connections. Most neural networks have some sort of training-rule.The training algorithm adjusts the weights on the basis of presented patterns. In other words, neural networks "learn" from examples.ANN's excel particularly at problems where pattern recognition is important and precise computational answers are not required. When ANNs inputs and/or outputs contain evolved parameters, their computational precision and extrapolation ability significantly increases and can even outperform more traditional modeling techniques.
See also: Boundary element modeling,Corrosion models, Knowledge based models, Mechanistic models, Pitting fatigue models, Risk based models
