Magazine article The National Public Accountant

Beyond Expert Systems: Neural Networks in Accounting

Magazine article The National Public Accountant

Beyond Expert Systems: Neural Networks in Accounting

Article excerpt

New information-processing technologies continue to impact accounting. Today, accountants routinely use expert systems. These are computer programs that capture knowledge and make recommendations much like human experts. Yet despite these advances, there has been no substitute for human knowledge and advice. Expert systems can only mimic the decision-making processes of human experts. Without human knowledge, an expert system would not exist.

Neural networks could change all this. Application of neural network technology can create new knowledge without the need for a human expert. Although neural networks have received little attention in accounting, many possible applications exist, particularly in those areas in which experts cannot clearly define or document their decision-making processes.

Neural networks, based on the fundamental processing unit of the brain - the neuron - simulate intellectual processes by linking inputs to outputs using a series of mathematical functions. Unlike expert systems that process captured knowledge, neural networks actually create knowledge. They reach beyond expert systems in their unique ability to handle incomplete, imprecise or partially incorrect data. They can be used in areas where knowledge is not available and facts and boundaries are unclear. Moreover, neural networks are adaptable and can be used under dynamic conditions, while expert systems are best reserved for static circumstances.

Neural Networks Versus Expert Systems

Expert systems and neural networks are both artificial intelligence applications. Today, users apply expert systems to problems that require years of special education and training. To develop an expert system, designers must understand the problem-solving process and "program" the process into the system. On the other hand, neural networks are not programmed. Instead, their developers train them by employing a series of examples and results. By matching input conditions to predetermined output results, the system derives mathematical functions that link the two.

A comparison of expert systems and neural networks provides insights into their design, development and operation. These applications differ with respect to their environment, function and structure. Exhibit 1 summarizes these differences.


Different system environments pose different problems that, in turn, require different solutions. The system environment of expert systems and neural networks may be compared on the basis of differences in problem domain, logic and reasoning.

Problem Domain

The domain defines the scope of system operation. Whereas expert systems have a narrow domain, the domain of a neural network is virtually limitless.

The rules defined by the experts when constructing the expert system constrain the operation of the system. System scope is limited to conditions addressed by the rules. Conditions not defined in the expert system's rules can not be processed by the expert system.

On the other hand, predefined rules do not limit the operation of neural networks. Rules in a neural network may be continuously modified, recognizing changing input or output conditions. Since the system can create new knowledge in response to changes in the environment, it has no specific domain. System inputs and outputs dynamically shape the operation of the neural network.

Problem Logic

Expert systems and neural networks also use different types of logic. Dependent on pre-defined rules, expert systems use rigid logic. Devoid of flexibility, an expert system applies a predetermined set of conditions to input data. All conditions and associated actions must be defined by the system or it will fail.

Conversely, neural networks use fuzzy logic. No clearly defined, predetermined set of conditions and actions constrain neural network operations. Instead, neural network operations are flexible. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.