Computational & Technology Resources
an online resource for computational,
engineering & technology publications 

CivilComp Proceedings
ISSN 17593433 CCP: 82
PROCEEDINGS OF THE EIGHTH INTERNATIONAL CONFERENCE ON THE APPLICATION OF ARTIFICIAL INTELLIGENCE TO CIVIL, STRUCTURAL AND ENVIRONMENTAL ENGINEERING Edited by: B.H.V. Topping
Paper 39
Life Cycle CostOriented Optimization of Steel Frames: A Neural Network Approach S.S. Abdelatif Hassanien and N. Shrive
Civil Engineering Department, University of Calgary, Alberta, Canada S.S. Abdelatif Hassanien, N. Shrive, "Life Cycle CostOriented Optimization of Steel Frames: A Neural Network Approach", in B.H.V. Topping, (Editor), "Proceedings of the Eighth International Conference on the Application of Artificial Intelligence to Civil, Structural and Environmental Engineering", CivilComp Press, Stirlingshire, UK, Paper 39, 2005. doi:10.4203/ccp.82.39
Keywords: structural reliability, life cycle cost, bilevel optimization, first and second order reliability method (FORM/SORM), Monte Carlo simulation, neural networks, finite element method.
Summary
In this paper a robust and efficient methodology is presented for solving life
cycleoriented optimization problems. The proposed methodology combines structural
reliability, optimization and neural networks intelligently. Over the last decades
optimization techniques and structural reliability analysis have stimulated the
probabilistic optimum design of structures. Despite of the advances in this field,
serious computational obstacles arise when treating practical problems. Life cycle
costoriented optimization (LCCOO) is simply minimizing the aggregation of all cost
components. Construction cost, annual maintenance and operation costs, repair costs and
expected damage and failure costs (economic losses, deaths and injuries) are examples
for different life cycle objectives. Here, two cost components are considered namely;
initial (construction) cost and the expected failure cost. The latter cost includes the
probability of failure. The available methods for estimating the probability of failure can
be roughly classified into two groups, which can be labelled as gradientbased and
simulationbased methods. Gradientbased methods such as first order/second order
reliability techniques (FORM/SORM) are optimizationbased techniques which turn the
LCCOO into a bilevel optimization problem. Simulationbased techniques such as
Monte Carlo simulation hinge upon the creation of a synthetic set of response samples.
Such techniques are timeconsuming procedures. Few methods are presented to solve
such probabilistic bilevel optimization issues [1,2,3] but the problem is still under
investigation.
For most realistic structures or systems, the response has to be computed through numerical procedures such as finite element analysis. This brings another complexity to the LCCOO because the performance function (limit state) is not available as an explicit, closedform function of the input variables. A two stage procedure for obtaining the optimum design is presented. Two neural networks are trained: one for the implicit limit state and the other to find the relation between the probability of failure and the vector of the design parameters. The use of neural networks is motivated by the approximate concept inherent in reliability analysis and the time consuming resulted from using the other documented methods. The proposed methodology utilizes neural networks intelligently to reduce the number of iterations required to build an approximation for the implicit limit state function under consideration. The paper discusses the need for having a design of experiments scheme for the vector of design parameters thorough an example. Another example tests the application of the approach on hand for the design of steel frames under static loads. The approach shows competitive results to those extracted from the onelevel technique and the second order response surface. It is expected that this methodology will be adopted with respect of applications of Level IV design codes. References
purchase the fulltext of this paper (price £20)
go to the previous paper 
