Expert system

from Wikipedia, the free encyclopedia

An expert system ( XPS or ES ) is a computer program that can support people in solving more complex problems like an expert by deriving recommendations for action from a knowledge base. So-called if-then relationships can be used to represent human knowledge (relationships in the world) in a way that computers can understand ( knowledge base ). An expert system contains the functionality to create and improve the knowledge base (knowledge acquisition component), to process it (problem-solving component) and to make it understandable to the user (explanatory component). Expert systems are a sub-area of artificial intelligence . Examples are systems to support medical diagnoses or to analyze scientific data. The first work on the corresponding software took place in the 1960s. Expert systems have also been used commercially since the 1980s.

Development history

The emergence of expert systems went hand in hand with the failure of another research goal in artificial intelligence, which is often referred to as the general problem solver . If one had initially tried to use general problem-solving approaches to arrive at a system that was supposed to generate solutions independently of the respective problem area, it soon found out that such a general problem solver could not be implemented and only achieved poor results for numerous questions. A larger knowledge base for generating solutions was necessary, especially for questions in special application domains. Expert systems are systems that are based on such a knowledge base, usually maintained by experts. In doing so, however, they by no means merely reproduce the content of the knowledge base, but are in a position to arrive at further conclusions on their basis . The quality of an expert system can be measured by the extent to which the system is able to draw conclusions and how error-free it is.

Realization principle

Very different models can be used to represent knowledge as well as to draw conclusions:

  • Case-based systems start from a case database which describes specific problems in their context, including a solution that has been implemented. For a given case, the system tries to find a comparable and as similar as possible case in its case database and to transfer its solution to the current case. The concept of the similarity of cases represents the key problem of such systems. A typical example of a case is a patient with his disease symptoms and the diagnostic measurement results. The solution sought here would be a correct diagnosis.
  • Rule-based systems or Business Rule Management Systems (BRMS) are not based on specific case descriptions, but on rules of the type "If A, then B". In contrast to cases, such rules are more general laws from which conclusions can be drawn for specific situations. Rules or business rules usually have to be entered directly into the system by human experts.
  • Another approach, which can be used in particular for classification problems, consists in systems that are independently capable of learning processes by means of decision trees . This is a form of inductive learning based on a sample set. An example can consist of a number of attributes (of an object, e.g. a patient) and their specific characteristics. When processing such examples, the system runs through a path (see also search tree ): The individual attributes are nodes, the possible characteristics that originate from them are edges. The system always follows the edge that applies in the present example, continues this process attribute by attribute and finally arrives at an end node (leaf). This finally indicates the class to which the described object is to be assigned. When building decision trees, the goal is to achieve the best possible classification results with the smallest possible trees. The difficulty here is the choice of attributes.

Knowledge base

In an expert system or knowledge based system that is knowledge base ( engl. Knowledgebase ) of the area of the system that includes the knowledge in any form of representation. The knowledge base is supplemented by an inference machine , i.e. software that can be used to operate on the knowledge base.


There is a need for expert system support wherever there is a lack of experts or the processing capacity of human experts is overwhelmed due to the complexity of the problem and the abundance of data material that arises. The application effect of expert systems is proportional to the complexity of the problem and the level gradient between an expert and the actual user. This difference in level is more serious the more complex and diffuse the problem area. The latter, in turn, is stronger the more inhomogeneously the area-relevant knowledge is structured and the less the area is formally penetrated, but rather is dominated by empirical knowledge.

Task classes and known expert systems

Typical task classes for expert systems are (the names of some of the implemented expert systems in brackets):

Data interpretation
Analysis of data with the aim of assigning them to objects or phenomena, especially signal understanding.
Examples : Recognition of acoustic language ( HEARSAY ), identification of chemical structures based on mass spectrometer data ( DENDRAL ), geological exploration ( PROSPECTOR ), protein structure determination from X-ray data , oil drilling, military reconnaissance, submarine location ( SEPS , STAMMER ).
Interpretation of data with triggering of actions depending on the result.
Examples : production assurance , monitoring of patients in the " iron lung " (VM), monitoring of a nuclear reactor ( REACTOR ).
Interpretation of data with a strong explanatory component.
Examples : diverse in medicine, for example bacterial infections ( MYCIN ), rheumatology, internal medicine ( INTERNIST ), plant diseases; also for the determination and localization of errors in technical systems.
Actions to correct faulty system states and eliminate the causes (often coupled with diagnosis).
Examples : see diagnosis, fault diagnosis in car transmissions (DEX), fault location and maintenance in telephone networks (ACE), automatic weaning of ventilated patients in intensive care medicine ( SmartCare / PS ), drug therapy safety ( CPOE , CDS ).
Generating and evaluating sequences of actions to achieve target states:
Examples : experimental planning of molecular genetic experiments ( MOLGEN ), chemical synthesis ( SECS ), financial planning ( ROME ), production planning ( ISIS), control of flight operations on aircraft carriers (CAT), actions of autonomous robots ( NOAH ), e.g. Martian robots.
Description of structures that meet specified requirements.
Examples : for circuit design (SYN, DAA), computer configuration ( R1 / XCON ), chemical compounds ( SYNCHEM ), configuration of operating systems for Siemens computers ( SICONFEX ).
Prediction and evaluation of achievable states of time-variant systems.
Examples : Assessment of earthquake effects ( SPERIL ), earthquake forecast , flood forecast , environmental development ( ORBI ).

Disadvantages in use

Expert systems can become counterproductive for solving a problem if users completely rely on them without intelligent support or if there is no constant intelligent search for alternative solutions. Because every expert system only has a limited amount of data, it is usually only fed data from the immediate vicinity of the problem. This creates the risk of missing important fundamental changes, only offering conservative solutions or explanations. The expert system cannot question the specified parameters, the entire system (see closed world assumption ). Inventions, innovations or the like require a creative combination of the problem with other - for example unrelated to the subject - knowledge (e.g. that a chocolate bar slips unnoticed into a petrol tank is not a programmable value for the petrol station expert system , which is why this case is not conceivable ) .

If expert systems are automated, disastrous effects can threaten in some areas of application, for example in the case of automated military actions that are not intelligently supervised.

There is a widespread view that Black Monday 1987 was partly caused or reinforced by the momentum of many computer traders who react very similarly .

See also


  • CLIPS : software tool for creating expert systems.
  • JESS : Rule system for the Java platform - successor and extension of the CLIPS rule system.
  • Prolog : a logic programming language for creating expert systems.



  1. P. Mertens et al.: Fundamentals of business informatics. 5th edition. Springer Verlag, Berlin 1998, ISBN 3-540-63752-4 , p. 53.
  2. FHC Naber: Fast solver for the three-factor Heston –– Hull – White problem. 2006, p. 11. (
  3. ^ Gary H. Jones, Beth H. Jones, Philip Little: Reputation as reservoir: The Value of Corporate Goodwill as a Buffer Against Loss in Times of Economic Crisis. 1999 ?, p. 2. (
  4. Frank Westerhoff: Bubbles and crashes: optimism trend extrapolation and panic. S. 5. ( ( Memento from June 27, 2007 in the Internet Archive ))