This thesis presents the optimization of neural network architectures using genetic algorithms. A key challenge in this area is the simultaneous determination of both network architecture and training hyperparameters, which in traditional approaches is often time-consuming and inefficient. To address this, we developed a chromosome representation that encodes hyperparameters and hidden layer parameters in a unified way, together with procedures for selection, mutation, and multiple crossover strategies. The approach incorporates a fitness function that balances predictive performance with a penalty for model complexity, as well as a progressive learning schedule. Based on these components, we designed a comprehensive algorithm for the automated search of suitable architectures and hyperparameters and compared it against the NEAT algorithm. The proposed method was evaluated on six benchmark datasets from the UCI repository, demonstrating competitive predictive performance, though at the cost of higher complexity and computational demand compared to NEAT, which produced simpler and faster models.
|
