Abstract

Bridges constitute a crucial component of urban highways due to the complexity and uncertain nature of their construction process. Simulation is an alternative method of analyzing and planning the construction processes, especially the ones with repetitive and cyclic nature, and it helps managers to make appropriate decisions. Furthermore, there is an inverse relationship between the cost and time of a project and finding a proper trade-off between these two key elements using optimization methods is important. Thus, the integration of simulation models with optimization techniques leads to an advancement in the decision making process. In addition, the large number of resources required in complex and large scale bridge construction projects results in a very large search space. Therefore, there is a need for using parallel computing in order to reduce the computational time of the simulation-based optimization. Most of the construction simulation tools need an integration platform to be combined with optimization techniques. Also, these simulation tools are not usually compatible with Linux environment which is used in most of the massive parallel computing systems or clusters.
In this research, an integrated simulation-based optimization framework is proposed within one platform to alleviate those limitations. A master-slave (or global) parallel Genetic Algorithm (GA) is used as a parallel computing technique to decrease the computation time and to efficiently use the full capacity of the computer. In addition, sensitivity analysis is applied to identify the promising configuration for GA and analyzing the impact of GA parameters on the overall performance of the specific simulation-based optimization problem used in this research. Finally, a case study is implemented and tested on a server machine as well as a cluster to explore the feasibility of the proposed approach.
The results of this research showed better performance of the proposed framework in comparison with other GA optimization techniques from the points of view of the quality of the optimum solutions and the computation time. Also, acceptable improvements in the computation time were achieved for both deterministic and probabilistic simulation models using master-salve parallel paradigm (8.32 and 20.3 times speedups were achieved using 12 cores, respectively). Moreover, performing the proposed framework on multiple nodes using a cluster system led to 31% saving on the computation time on average. Furthermore, the GA was tuned using sensitivity analyses which resulted in the best parameters (500 generations, population size of 200 and 0.7 as the crossover probability).