Prediction of uplift capacity of suction caissons using a neuro-genetic network

Suction caissons are frequently used for the anchorage of large offshore structures. The uplift capacity of the suction caissons is a critical issue that needs to be predicted reliably. A neuro-genetic model has been employed for this purpose. The neuro-genetic model uses the multilayer feed forward neural network (NN) as its host architecture and employs genetic algorithms to determine its weights. In comparison to the application of a conventional NN model [49] for the uplift capacity prediction problem, the application of a hybrid model such as the neuro-genetic network appears attractive. The conventional NN model is sensitive to training parameters and initial conditions and calls for a longer training of the network. Also it is not free of the inherent problem of settling for the local minimum in the neighborhood of the initial solution. In contrast, the hybrid model is much less sensitive to training parameters and initial conditions and inherently looks for a global optimum in a complex search space, which may be multimodal or non-differentiable, with a modest amount of training. The performance of the neuro-genetic model has been studied in detail over specific data sets pertaining to suction caissons, gathered from 12 independent studies [49] and compared with the predictions made by NN and finite element method models.