R. HAMID, KHAIRULLAH YUSUF, ABDUL KHALIM ABDUL RASHID Department of Civil and Structural Engineering Faculty of Engineering and Built Environment Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor MALAYSIA. roszilah@eng.ukm.my khairull@eng.ukm.my The paper presents result from experiments on network architecture and transfer functions configuration in the feed1forward neural networks (FFNNs) applied to bridge condition rating approximation. Trial and error approach is done on three layers feed1forward neural network by varying the number of neurons in hidden layer. Levenberg1Marquardt training algorithm () and transfer function are applied in FFNN to investigate the best configuration to be used for bridge condition rating. Mean square error ( ) and correlation coefficient () are used to measure the network performance. The results indicated that the configuration of FFNN with thirty1one neurons in hidden layer using transfer function in output layer have produced the best and than other configurations. 1 Feed1forward neural networks, bridge condition rating, transfer function. Bridge condition rating is the most important part in the bridge management system. This is because record of bridge condition rating data can affect approximately 60 % of the bridge management system analysis modules [1]. For bridge management system needs, usually bridge authorities will apply visual inspection as the first step of any condition assessment procedure unless a structure cannot be visually assessed. Through visual inspection, bridge inspectors evaluate the condition of a bridge using their personal experience and following guidelines such as those found in some inspector’s manual. These evaluation incorporates many parameters and human judgements, although are very valuable, that can sometimes be vague [2]. It would be ideal to have physical structural tests on each bridge structure in the system, but it would be impractical and economically prohibitive to implement them, given the large number of bridges to be inspected within a given period. As an alternative, the collective judgment of the inspectors can be used to develop unified, coherent bridge inspection procedures [3]. Thus a procedure such as artificial neural network would be useful to handle this uncertainty, imprecision, and subjective judgment. Artificial neural networks (ANNs) are widely used as attractive alternative to handle complex and non1linear systems that are difficult to model using conventional modelling techniques such as mathematical modelling. It has been widely applied in engineering, science, medical, economic, physical and environmental application. The most common applications are function approximation, pattern classification, clustering and forecasting [41 5]. Various forms of artificial neural networks (i.e. feed1forward neural network, recurrent networks, radial basis functions, wavelet neural network, hopfield network, etc.) have been applied in various disciplines. However, in the context of function approximation like bridge condition rating, the feed1forward neural network (FFNN) is generally chosen as the network architecture [6] and back1 propagation (BP) as the learning algorithm [718]. Generally, the successful application of artificial neural networks for purpose of prediction and modelling in science and engineering domains is tremendously affected by the consequent main factors: network form, network architecture, training algorithms, transfer functions, input selection, neural network weight, momentum rate, number of iterations, and data set partitioning ratio [9]. The literature review has been made and there is no exact available formula to decide how many hidden layers, how many numbers of neuron in each hidden layer, which training algorithm and transfer function can solve a given problem. A trial Latest Trends on Engineering Mechanics, Structures, Engineering Geology ISSN: 1792-4294 408 ISBN: 978-960-474-203-5