These tools were generally closed source and hard or impossible to extend and could be difficult to operate due to limited accompanying documentation. Machine learning software has been primarily comprised of proprietary software like MATLAB® with the Neural Networks Toolbox and Wolfram Mathematica or independent university projects like the Stuttgart Neural Network Simulator (SNNS). This decade of the 2000s introduces a shift in tooling, which is a direct contributor to the recent increase in adoption and research of both shallow and deep machine learning research. Jesper Sören Dramsch, in Advances in Geophysics, 2020 2.1 Modern machine learning tools These preprocessing functions must be reapplied to the inputs and outputs of the implemented network. The training inputs to the neural network are first preprocessed by the Matlab mapstd or mapminmax functions and the training outputs of the neural network are preprocessed using the mapminmax function. Training was undertaken using the Bayesian regularization training function of the Matlab toolbox using the sum of squared error performance measure. The number of input delay states was set to 12 hours, and output delay states set to 1 hour. As more target measurements are included the size of the hidden layer increases. For the NARX architecture, the number of past time samples, or delay states, considered also impacts the performance of the network.įor this work, a network structure of one hidden layer with 10 neurons was found to provide reasonable training time and acceptable performance when training for an individual target measurement. In general, the larger the network the more complex the function that can be approximated, however if the network is too large it may not generalize well to new data. The numbers of neurons in the input layer, hidden layer and output layer is difficult to determine ( Sbarbaro et al. Lab samples are used to correct the neural network feedback.

The resulting, trained series-parallel network is then converted to a parallel architecture. A series-parallel network architecture is used with a static back-propagation training algorithm to reduce training time. The nonlinear autoregressive network with exogenous inputs (NARX) neural network structure is employed. The Mathworks Matlab Neural Network toolbox is utilized to build and train a suitable neural network. Ali Nooraii, in Computer Aided Chemical Engineering, 2011 3.2 Neural Network Estimator structure and training

Matlab 2018b ut chemical engineering