A Hyperparameter Search Approach Using BRKGA for Deep Feedforward Neural Network

This paper proposed to use the Biased Random-Key Genetic Algorithm (BRKGA) to optimize the main hyperparameters of a Deep Feedforward Neural Network (DFNN). The proposed methodology was detailed and compared with an approach based on the Tabu Search algorithm known in the literature and used for the same purpose in six datasets. Both algorithms showed high-performance results, but BRKGA resulted in better solution quality, faster convergence with fewer iterations, and better performance locally. Finally, a statistical test was used to prove the statistical difference between the two algorithms, thus guaranteeing the preference of the BRKGA regarding the results.