Genetic Algorithm-Based Search Space Exploration to Generate Best Convolutional Neural Network

Despite the wide variety of applications and use cases that can be solved with the help of machine learning algorithms, researchers have yet to develop a general artificial intelligence solution that can solve different problems. A single machine learning model can provide at least or above-average human-level performance in multiple domains without re-training for each domain separately. Currently, machine learning solutions train multiple models for each domain and combine their outputs to generate high-performance in multi-domain results. While this multi-model approach is achievable, designing and training each model and fine-tuning hyperparameters is complex and time-consuming. However, automated machine learning approaches provide an algorithmic and probabilistic approach to training models with the best-performing hyperparameters. This paper, after taking inspiration from gene theory of adaptation, presents a Novel Inverse Shuffle crossover using the Genetic algorithm method that generates high-performance convolutional neural network architectures for search space exploration (GA-NISC). New integrations enable the proposed method to explore the hyperparameter search space more thoroughly before converging to the best-performing model architecture as compared to its peers. The experiments are performed on a benchmark German traffic signal classification dataset and the CIFAR-10 dataset. Experimental results revealed that the proposed method converged faster in contrast to other neural architectural search methods while reducing the training time and overall computations considerably. Besides, it also yield high fitness scores up to 92% and 98% respectively under the best generated architecture. Furthermore, a composite fitness metric was calculated by integrating Validation Accuracy, F1-Score, and Regularization into the cross-entropy loss metric to meaningfully improve model generalization, especially on imbalanced data.