posted on 2021-05-21, 09:40authored byJames VanderVeen
Machine learning models can contain many layers and branches. Each branch and layer, contain individual variables, know as hyperparameters, that require manual tuning. For instance, the genetic algorithm designed by Unit Amin [2] was designed to mimic the reproductive process of living organisms. The genetic algorithm and the Artificial Neural Network (ANN) training processes contain inherent randomness that reduces the replicability of results. Combined with the sheer magnitude of hyperparameter permutations, confidence that model has arrived at the best solution may be low. The algorithm designed for this thesis was designed to isolate portions of a complex ANN model and generate results showing the effect each hyperparameter has on the performance of the model as a whole. The results of this thesis show that the algorithm
effectively generates data correlating model performance to hyperparameter selection. This is evident in section 3.1, and 3.2, where it is shown that using the sigmoid activation function with CNN layers, regardless of the number of filters, or hyperparameters used in the subsequent LSTM layers, produces superior RMSE scores. Section 3.2 also reveals that the model does not improve in performance as the number of CNN and LSTM layers are added to the model. Finally, the results in section 3.3 show that the rmsprop optimizer generates superior results regardless of the hyperparameters used in the rest of the model.