Input Variable Selection of Artificial Neural Network for Prediction of Stresses in Extended Shear Tab Connection

A connection with an extended shear tab (EST) must withstand both the shear and the moment transmitted from the supported beam. EST connection is widely used in beam-to-column flange and beam-to-girder web arrangements and it has shown to be both inexpensive and simple to install. The current paper presents Artificial Neural Network (ANN) to predict stress values in component parts of EST connections. The selection of the inputs is an important part of the development of an ANN. Three different input selection techniques, including Trial and Error, Correlation Analysis and Average Mutual Information (AMI) are used to determine the final inputs for the models. Trial and error is an efficient approach of selection of the inputs but its tedious method. Recently, correlation analysis between input variables and outcome variables has been used to choose inputs. However, linear correlation analysis makes it inappropriate for application in present case, as the issue at hand is non-linear. The AMI is a non-linear technique to measure the relationship between two variables. In the present study, inputs are chosen using all three aforementioned techniques and feed-forward ANN models are developed to predict stresses in each components of EST connection. ANN model shows an acceptable performance of correlation coefficient i.e., greater than 0.8 and low error measures i.e., RMSE less than 0.08 and MAE less than 0.07. To support this scatter plots also depicts the same results. The results show that, out of the three techniques, ANNs developed model using the AMI method produces the best outcomes.