On the optimization of articial neural networks for application to the approximation of chemical systems

An arti cial neural network (ANN) is a computational model for storing and retrievingacquired knowledge. ANNs consist of dense interconnected computing units that are sim-ple models for complex neurons in biological systems. The knowledge is acquired during alearning process and is stored in the synaptic weights of the inter-nodal connections. Themain advantage of neural networks is their ability to represent complex input/outputrelationships. They are well suited for use in data classi cation, function approximation,and signal processing, among others.The performance, or tness, of an ANN is often measured according to an error betweentarget and actual output, training time, complexity of the ANN, or in terms of otherproperties important for the user. Although the elemental building blocks of a neuralnetwork, i.e., neurons, nodal connections, and the transfer functions of nodes are inthemselves relatively simple, the various combinations can result in di erent topologieswith similar or vastly di erent tness characteristics. Therefore, the a priori design of aneural network with near-optimal tness is not a trivial task and is usually guided byheuristics or trial-and-error. The architecture or topological structure of an ANN can becharacterized by the arrangement of the layers and neurons, the nodal connectivity, andthe nodal transfer functions. In this work, the class of multi-layer perceptrons (MLPs) isconsidered, which consists of an input layer with N