2024年3月29日发(作者：联想a7000一体机拆机图解)

Ultsch, A. & Korus, D. „Integration of Neural Networks with Knowledge-Based Systems“ Proc. IEEE Int. Conf. Neural Networks, Perth/

Australia, 1995.

Integration of Neural Networks with

Knowledge-Based Systems

Alfred Ultsch, Dieter Korus

Department of Mathematics/Informatics; University of Marburg

Hans-Meerwein-Straße/Lahnberge; D-35032 Marburg; F. R. Germany

email: ultsch or korus@

/~wina/

ABSTRACT

Existing prejudices of some Artificial Intelligence researchers against neural networks are hard to

break. One of their most important argument is that neural networks are not able to explain their

decisions. Further they claim that neural networks are not able so solve the variable binding pro-

blem for unification. We show in this paper that neural networks and knowledge-based systems

must not be competitive, but are capable to complete each other. The disadvantages of the one

paradigm are the advantages of the other and vice versa. We show several ways to integrate both

paradigms in the areas of explorative data analysis, knowledge acquisition, introspection, and

unification. Our approach to such hybrid systems has been prooved in real world applications.

1. Introduction

The successful application of knowledge-based systems in different areas as diagnosis, construction and plan-

ning shows the usefulness of a symbolic knowledge representation. However, this representation implies pro-

blems in processing data from natural processes. Normally such data are results of measurements and have

therefore no straightforward kind of symbolic representation [1]. Knowledge-based systems often fall short in

handling inconsistent and noisy data. It is also difficult to formalize knowledge in such domains where ´a priori´

rules are unknown. Often the performance in ´learning from examples´ and ´dealing with untypical situations´

(graceful degradation) is insufficient. The rules used by conventional expert systems are sait to be able to repre-

sent complex concepts only approximately [4]. In such complex systems inconsistent and context-dependent

rules (cases) may result in unacceptable errors. In addition, it is almost impossible for experts to describe their

knowledge, which they acquired from many examples by experience, entirely in symbolic form [6].

State-of-the-art knowledge-based system technology is based on symbolic processing. Acknowledged shortco-

ming of current computational techniques is their brittleness, often arising from the inability of first order logic

to capture adequately the dynamics of a changing and incompletely known environment. An important property

of knowledge stored in symbolic form is that it can be interpreted and communicated to experts. The limits of

such an approach, however, become quite evident when sensor data or measurement data, for example from

physical processes, are handled. Inconsistent data frequently force symbolic systems into an undefined state.

Another heavy problem in knowledge-based system design is the acquisition of knowledge. It is well known that

it is almost impossible for an expert to describe his domain specific knowledge entirely in form of rules or other

knowledge representation schemes. In addition, it is very difficult or even impossible to describe expertise

acquired by experience.

Neural networks claim to avoid most of the disadvantages of knowledge-based systems described above. These

systems which rely on a distributed knowledge representation are able to develop a concise representation of

complex concepts. It is possible to learn knowledge from experience directly [4]. Characteristic attributes of

connectionist systems are the ability of generalization and graceful degradation. E.g. they are able to process

inconsistent and noisy data. In addition, neural networks compute the most plausible output to each input.

Neural networks, however, also have their disadvantages. It is difficult to provide an explanation of the beha-

viour of the neural network because of the distributed knowledge representation. Therefore expertise learned by

neural networks is not available in a form that is intelegible for human beings as well as for knowledge-based

Ultsch, A. & Korus, D. „Integration of Neural Networks with Knowledge-Based Systems“ Proc. IEEE Int. Conf. Neural Networks, Perth/

Australia, 1995.

systems. It seems to be difficult to describe or to interpret this kind of information. In knowledge-based systems

on the other hand it is easy to describe and to verify the underlying concepts.

2. Integration of Neural Networks with Knowledge-Based Systems

Indications are that neural networks provide fault-tolerance and noise resistance. They adapt to unstable and lar-

gely unknown environments as well. Their weakness lies in a reliance on data-intensive training algorithms,

with little opportunity to integrate available, discrete knowledge. At present, neural networks are relatively suc-

cessfull in applications dealing with subsymbolic raw data; in particular, if the data is noisy or inconsistent. Such

subsymbolic level processing seems to be appropriate for dealing with perceptions tasks and perhaps even with

tasks that call for combined perception and cognition. Neural networks are able to learn structures of an input set

without using a priori information. Unfortunately they cannot explain their behavior because a distributed repre-

sentation of the knowledge is used. They only can tell about the knowledge by showing responses to a given

input.

Both approaches, knowledge-based systems and neural networks, of modelling brain-like information proces-

sing are complementary in the sense that traditional knowledge-based systems are a top-down approach starting

from high-level cognitive functions whereas neural networks are a bottom-up approach on a biophysical basis of

neurons and synapses. It is a matter of fact that the symbolic as well as the subsymbolic aspects of information

processing are essential to systems dealing with real world tasks. Integrating neural networks and knowledge-

based systems is certainly a challenging task [10]. Beside these general considerations several specific tasks

have to be solved. The most important are - without claiming on completeness:

Structure Detection by Collective Behavior: In real world people have continuously to do with raw and sub-

symbolic data which is characterized by the property that one single element does not have a meaning (interpre-

tation) of itself alone. The question is, how to transform the subsymbolic data into a symbolic form.

Unsupervised learning neural networks can adapt to structures inherent in the data. They exhibit the property to

produce their structure during learning by the integration (overlay) of many case data. But they have the disad-

vantage that they cannot be interpreted by looking at the activity or weights of single neurons. Because of this

we need tools to detect the structure in large neural networks.

Integrated Knowledge Acquisition: Knowledge acquisition is one of the biggest problems in artificial intelli-

gence. A knowledge-based system may therefore not be able to diagnose a case which an expert is able to. The

question is, how to extract experience from a set of examples for the use of knowledge-based systems. Under

Integrated Knowledge Acquisition we understand subsymbolic approaches, i.e. the usage of neural networks, to

gain symbolic knowledge. Neural networks can easily process subsymbolic raw data by handling noisy and

inconsistent data. An intrinsic property of neural networks is, however, that no high level knowledge can be

identified in the trained neural network. The central problem for Integrated Knowledge Acquisition is therefore

how to transform whatever a neural network has learned into a symbolic form.

Introspection: Under introspection we understand methods and techniques whereby a knowledge-based system

observes its own behaviour and improves its performance. This approach can be realized using neural networks

that observe the sequence of steps an expert system takes in the derivation of a conclusion. This is often called

control knowledge. When the observed behaviour of the expert system is appropriately encoded, a neural net-

work can learn how to avoid missleading paths and how to arrive faster at its conclusions.

Unification: One type of integrated reasoning is the realization of an important part of the reasoning process, the

unification, using neural networks. Unification pays a central role in logic programming (e.g. in the language

Prolog) and is also a central feature for the implementation of many knowledge-based systems. The idea of this

approach is to realize the matching and unification part of the reasoning process in a suitable neural network.

3. Structure Detection by Collective Behavior

One of the neural network types we use for representing subsymbolic raw data in large distributed neural net-

works are the Self-Organizing Feature Maps (SOFM) by Kohonen [5]. It has the ability to map a high-dimen-

sional feature space onto a usually two-dimensional grid of neurons. The important feature of this mapping is

that adjacent points in the data space are mapped onto adjacent neurons in the grid by conserving the distribution