Expert systems with applications

Solo disponible en BuenasTareas
  • Páginas : 19 (4688 palabras )
  • Descarga(s) : 0
  • Publicado : 5 de febrero de 2012
Leer documento completo
Vista previa del texto
Expert Systems with Applications
PERGAMON
Expert Systems with Applications 21 (2001) 31±36 www.elsevier.com/locate/eswa

Neural networks as a tool for developing and validating business heuristics
Steven Walczak*
University of Colorado at Denver, College of Business and Administration, Campus Box 165, PO Box 173364, Denver, CO 80217-3364, USA

Abstract The increasing availability ofinformation via the Internet and world-wide-web has made business decision-making more complex. Heuristic rules and methods are frequently used in complex domains to facilitate the decision-making process by reducing the amount of information that is required for decision-making. Heuristic rules may `go out of date' as new information and new business methods are developed. Neural networks provide afast and ef®cient means for evaluating the utility of existing heuristics. Two case studies are presented that demonstrate the use of neural networks for developing new heuristic rules or for refuting existing heuristic rules. Validation or adaptation of non-valid heuristics improves the quality of resulting business decisions. q 2001 Published by Elsevier Science Ltd.
Keywords: Neural networks;Heuristics; Model selection; Empirical evaluation

1. Introduction As the world-wide-web continues to grow and more and more information is made available to managers, the decision-making process is becoming more complex. Solutions to business problems, especially in complex non-linear domains, are often approximated through the use of heuristics (Barr & Feigenbaum, 1981). Heuristic methodsserve to reduce the quantity of information that must be evaluated in order to reach a near-optimal solution. Although heuristic decision-making methods do not guarantee an optimal solution, they do approximate the optimal solution and the tradeoff between optimal and near-optimal solutions is often desirable due to the time and information cost savings afforded through the use of heuristics. Sinceheuristics are typically based on induction, two problems face business managers utilizing heuristic decision strategies. First is the validity of the heuristic. Induction uses empirical evidence to formulate the heuristic rules. The exponential growth of the availability of information implies that new evidence may refute a heuristic. However, selecting the right information from the mountainof available data is problematic. Second, is the possibility that even though the heuristic may have been valid at one point, the dynamic state of our global economy may produce needed changes to the heuristic methods (e.g., changing from a hierarchical management structure to a matrix organization). Arti®cial intelligence has provided a number of meth* Tel.: 11-303-556-6777. E-mail address:swalczak@carbon.cudenver.edu (S. Walczak). 0957-4174/01/$ - see front matter q 2001 Published by Elsevier Science Ltd. PII: S 0957-417 4(01)00024-0

odologies for addressing complex problems and utilizing heuristic methods to produce solutions to these problems. Among these methodologies are: neural networks, expert systems, genetic programming, fuzzy systems, and others (e.g., arti®cial life).Current rule-based expert systems may implement hundreds or even thousands of heuristic rules. Validation of expert systems is a problematic issue and requires signi®cant development time to perform proper validation (Medsker & Liebowitz, 1994). Additionally, expert systems tend to be static in their use of heuristics and therefore do not facilitate the adaptation or refutation of current heuristics.Neural networks and genetic programming however, utilize automatic system adaptation (learning) to new empirical values and hence can promote the evaluation and adaptation of heuristic decision rules. This article discusses the utility of using neural networks for evaluating business decision heuristics. A very brief background on neural networks and the model selection research methodology is...
tracking img