Miriyala, S S
(2015)
Analysis of Artificial Neural Networks For Building Automated Surrogate Algorithms.
Masters thesis, Indian Institute of Technology Hyderabad.
Abstract
While attaining the objective of online optimization of complex chemical processes, the possibility of using the first principle based models is rarely an option, since such models demand large computational time. Surrogate models, which can emulate first principle based models, offer a credible solution to this problem, by ensuring faster optimization. Thus, the entire challenge of enabling
online optimization of complex models depends on construction of efficient surrogate models. Often, the surrogate building algorithms have certain parameters that are usually fixed based on some heuristic, thereby inviting potential errors in building such surrogate models. This work aims at presenting an elaborate study on the effect of various parameters affecting the predictability of artificial neural networks viz.(a) architecture of ANN, (b) sample size required by the ANN, (c) maximum possible accuracy of prediction, (d) a robust sampling plan and (e) transfer function choice for node activation. The ANNs are then utilized as surrogates for a highly nonlinear industrial sintering process, the optimization of which is then realized nearly 7 times faster than the optimization study using the expensive phenomenological model.
Actions (login required)
 |
View Item |
Altmetric
Altmetric