Artificial neural networks to aid conceptual design

Activities related to the conceptual stage of the design process are based mainly on human intelligence, intuition, evolutionary ideas, and past successful experience. Uncertainties in the design information (as it is incomplete at this stage) and a lack of clarity in the design brief are distinctive characteristics of this stage of the design process. Although computers are used to model a variety of engineering activities, at present the main focuses of computer applications are areas with well-defined rules, such as sophisticated analysis, graphical and CAD applications. Activities related to the conceptual stage of the design process are generally untouched by computers. In the past, knowledge-based expert systems (KBESs) tried to model some of the activities of conceptual design. Owing to their restricted scope, the success of these systems was very limited. Artificial neural networks (ANNs) are applications of artificial intelligence (AI) which can imitate the activities of the human brain. Like human experts, ANNs are capable of learning and generalising from examples and experience to produce meaningful solutions to problems, even when input data contain errors or are incomplete. This makes ANNs a powerful tool for modelling some of the activities of the conceptual stage of the design process. Research has shown that genetic algorithms (GAs) are powerful tools that are used to assist the designer at the conceptual stage of the design process. Research at the University of Plymouth has shown that, by integrating ANNs with the GAs, it is possible to develop computer aided tools that are capable of assisting designers at the earlier stages of the design process. These tools will enable the designer to investigate a wider range of design possibilities in a very short time and to make an informed decision on the choice of final design. The paper presents the architecture of a back-propagation NN, commonly used in practice. Methodologies for selecting appropriate NN topology and data selection for the NN training are discussed. An example of a continuous reinforced concrete flanged beam for NN training is given, and the results of the NN learning are presented.