Artificial reasoning models (e.g., Bayesian networks) can support business
and organizational decision-making activities. By artificial reasoning models,
valuable expertise can be shared among your organization.
Computerized diagnostic decision support and expert systems are often modeled on the basis of models that can reason: what could be the causes of my observations? Which additional measurements should I do?
Examples of diagnostic decision support
- Medical diagnosis: The computer assists the physician in making a diagnosis, in particular to draw attention to alternatives that may be overlooked.
- Machine diagnosis: what is the cause of machine failure (abnormal sensor readings, wear)? Here a support system may be of use for maintenance engineers throughout the company.
- The concept is more general: given current reported observations, what is the state of my system. Another application is a decision support system for geologists that want to estimate the contents of a potential oil field given the borehole measurements. In addition, the question is if further measurement actions (which are expensive) are needed.
Modern systems are often based on Bayesian network technology. A Bayesian network is a model in which variables are
organized in a network of direct influences. The model is defined in a forward way (from cause to effect). However,
reasoning is in all directions. It includes forward reasoning (from cause to effect), backward reasoning (from effect to cause),
but also combinations (e.g. from a risk for a cause and an effect to the actual cause; another example is to find the most
informative measurements to refine a diagnosis). The modular architecture of Bayesian networks makes large scale expert systems possible.
In a Bayesian network, the direct influences are with a certain probability. In the medical example: flu does cause headache with a certain probability (which is less than 100%). Bayesian networks are so-called probabilistic models. This means that they can reason with uncertainty, e.g. due to incomplete information or incomplete knowledge. E.g., again in the medical example, there may be hundreds of tests, but for given patient the diagnosis should be made on the basis of only a few of them.
In other words, the set of variables that is observed may differ from case to case. For a probabilistic model, this is not a problem. This is in contrast to e.g. a functional approach, such as with neural networks, which are less appropriate for such applications.
Bayesian networks can combine domain knowledge and statistical information. In practice, they often rely mostly on domain expertise from specialists. Consequently, by these models valuable expertise can be shared among your organization. In addition, these models encourage a uniform and consistent approach within your organization.
Bayesian networks contain a full domain description. Therefore they can be used for a variety of applications, such as what-if analysis, or case simulation - e.g. for training and education purposes.
PETROPHYSICAL DECISION SUPPORT
For SHELL, we have built a Bayesian network for petrophysical decision support.
Read more about this taylor made product.
SMART Research BV has the tools and the expertise to build reasoning (e.g. bayesian network) applications.
More about realizing applications
For our Bayesian network applications, we have developed BayesBuilder. BayesBuilder consists of a stand-alone application for Bayesian network model development and a JAVA library to include a Bayesian network engine in a software application.
The stand-alone application is freely available for academic usage. Ask us for more information.