Hybrid modelling of multilayer perceptron ensembles for predicting the response of bolted lap joints

Resumen

In the early design phase of steel structures, the maximum capacity of the connections is one of the critical parameters that ought to be calculated first. Determining this force by means of analytical methods usually leads to overestimations, given its highly non-linear behaviour. The finite element (FE) method, a hard computing approach, represents one of the best alternatives to obtain more realistic connection responses, except for its high costs in terms of computation time. The present study proposes the use of a hybrid artificial intelligence system to accurately predict the response of a particular type of steel connection: the bolted lap joint. The proposed hybrid system is composed of hard and soft computing components. First, a set of 800 FE simulations of different joint configurations was conducted to generate the training and testing data sets for the development of the system. Secondly, a multilayer perceptron network ensemble model was trained and tested. Additionally, a procedure based on genetic algorithms was included to optimize simultaneously both the settings of the model and the number of input variables involved in the process. The optimized ensemble model is compared to other soft computing alternatives and shows higher generalization capacity when it deals with testing data. Finally, the results support the use of this hybrid system to create prediction models with similar performance to the FE method in terms of accuracy but with a greatly reduced computational effort. © The Author 2015.