Construction of Optimal Artificial Neural Network Architectures for Application to Chemical Systems: Comparison of Generalized Pattern Search Method and Evolutionary Algorithm

摘要: Artificial neural networks (ANNs) are computational models of their biological counterparts. They consists densely interconnected computing units that work together to solve a specific problem. The information, which is acquired during learning process, stored in the synaptic weights internodal connections. main advantage ability represent complex functions and efficient storage information. ANNs frequently employed applications involving data classification, function approximation, signal processing (Haykin, 1994). topology an arrangement neurons, equipped with transfer weights, nodal Despite these simple topological elements, flexible neurons connections allows generation arbitrary complexity. resulting complexity, however, directly affects network performance. performance, or fitness, measure accuracy representing input-output relation. For instance, topologies only few provide limited flexibility functions. have therefore typically poor fitness. On other side, large new data, can lead generalizability extensive costs for training retrieval (Yao, 1999). Considering integration such large-scale simulations, from significant increase overall time. Because inherent complexity it apparent priori identification near-optimal performance challenging task, often guided by heuristics trial-and-error. design optimal be formulated as optimization choice method this problem determined properties ANN (Miller et al., 1989):