Optimization of Neural Network Training for Wine Quality Classification Using Simulated Annealing

Abstract

The backpropagation algorithm is the most used algorithm to train a neural network. However, a simulated annealing algorithm can do that work too. This paper shows the process and results of training neural network by applying simulated annealing algorithm. A Wine Quality Dataset is used to do the experiment. The experiments first extracted the data by feature selection and pre-processing of the dataset. By applying the Principal Component Analysis method, the features in the original data are extracted into a lower-dimensional space. The importance of features will increase significantly, and there is a positive effect on the training of neural networks. Then a variety of neural networks with different structures are constructed and trained with simulated annealing and back propagation respectively. More specifically, neural networks with two-hidden-layer fully-connected neural networks with two, three, and four hidden nodes in each layer are constructed to represent the different architecture of the network. Finally, their respective prediction results are compared to get a conclusion. This paper uses four parameters, accuracy, precision, recall and F1 score respectively, to evaluate the performance of the two target models, in addition to measure their performance in a more holistic way. As a result, the simulated annealing algorithm performs better than the backpropagation method in the context of wine quality classification.