A Design and Development Method for Artificial Neural Network Projects

MA-Thesis / Master von Stefan Vogt

Abstract:

In the 1980s research efforts and successes made artificial neural networks popular. Since the 1990s engineers have been using this foundation for problem solving. But artifiial neural network solutions for "real-world" problems are sometimes hard to find because of the complexity of the domain and because of the vast number of design attributes the engineer has to deal with.

This thesis provides a structured overview of attributes in the design process of artificial neural networks and reviews technical process models. Current development methods for artificial neural networks are then reviewed and critiqued. The thesis concludes with a new design and development method for artificial neural networks.

Table of Contents:

	List of figures	x
	List of tables	xi
	Introduction	1
1.	Design attributes in ANN	3
1.1	ANN models	4
1.1.1	Node level	7
1.1.2	Network level	9
1.1.3	Training level	9
1.2	Data and data representation	10
1.3	Global system design	12
1.4	Hardware and software implementation	13
1.5	Characteristics of ANNs	15
1.5.1	Advantages of ANNs	15
1.5.2	Limitations and concerns	16
2.	Technical process models and engineering methods	18
2.1	Why use an engineering method?	18
2.2	Evolutionary model of engineering discipline	20
2.3	Overview of technical process models	22
2.3.1	Taxonomy of technical process models	24
2.3.2	Prototyping	25
2.3.3	Incremental method	26
2.3.4	Strict contractual approach	26
2.3.5	Deciding on process models and methods	26
2.3.6	Examples of process models	27
2.3.7	Representation of process models	27
2.4	Quality criteria of process models	29
3.	Current engineering methods for ANNs	30
3.1	Why a special method for ANNs?	30
3.1.1	Are conventional engineering methodologies suitable for ANNs?	30
3.2	Methods for expert systems	31
3.3	System identication methods	35
3.4	Bailey and Thompson	37
3.4.1	Criticism	43
3.5	Medsker and Liebowitz	44
3.6	Jones and Franklin	45
3.7	Schalko	47
3.8	Karayiannis and Nicolaos	48
3.8.1	Criticism	49
3.9	Nelson and Illingworth	50
3.9.1	Criticism	51
3.10	Whittington and Spracklen	52
3.10.1	Criticism	56
3.11	Lawrence and Andriola	57
3.11.1	Criticism	58
3.12	General criticism of current methodologies	58
4.	Proposed design and development method	60
4.1	Development process	61
4.1.1	Requirement analysis	65
4.1.2	Specication	68
4.1.3	Data and domain analysis	70
4.1.4	Architectural design	76
4.1.5	Detailed ANN design	84
4.1.6	ANN implementation	92
4.1.7	Training	93
4.1.8	Monitoring training	94
4.1.9	ANN quality evaluation	95
4.1.10	ANN verication and validation	95
4.1.11	System link	96
4.1.12	System verication and validation	97
4.1.13	Delivery and production	97
4.1.14	Maintenance	97
4.2	Risk analysis	98
4.2.1	Engineers logbook	100
4.2.2	Review of stage results	101
4.3	Reuse	101
4.4	Integration of ANN development	104
4.5	Criticism	104
	Conclusion	107
	Appendix	109
	Bibliography	111