ABSTRACT
Ghaboussi and Joghataie 1995; Bani-Hani and Ghaboussi 1998a&b; Hung et al. 2000), two neural network models are used for active control of structures. A neural networkbased emulator is created to predict the structural responses and a neural network-based controller is created to determine the control forces required to satisfy the control criteria. Both neural network models need to be trained using structural response data. These published works are applied mostly to small two-dimensional frame structures. The approach of using a neurocontroller along with a neuroemulator is effective for small
structures with a few members. For large three-dimensional structures such as high-rising buildings, however, the
network size of the neurocontroller becomes increasingly large and the training of the
resulting network becomes prohibitively expensive. It should be pointed out that training
of a neuroemulator is based on a supervised learning algorithm such as the
backpropagation (BP) algorithm (Rumelhart et al. 1986) and the adaptive LM-LS
algorithm developed by the authors (Jiang and Adeli 2005a) and presented in Section
14.3. The training of a neurocontroller, on the other hand, is based on the trained
neuroemulator and some kind of unsupervised learning algorithm (Ghaboussi and
Joghataie 1995; Bani-Hani and Ghaboussi 1998b; Hung et al. 2000) because the magnitudes of control forces are problem-(structure-and earthquake-) dependent and
therefore unknown. Training of a neurocontroller requires a considerable amount of processing time, especially for large structures such as high-rising building structures.
