GA-Adam: A Hybrid Optimization Framework Balancing Exploration and Exploitation

This paper presents a hybrid optimization method combining the Genetic Algorithm (GA) and the Adaptive Moment Estimation (Adam) optimizer to balance global exploration and local exploitation.GA performs robust global search but converges slowly near local optima, while Adam achieves fast local convergence yet struggles with poor local minima.The proposed two-phase Genetic Algorithm-Adaptive Moment Estimation (GA-Adam) approach first evolves candidate solutions using GA, then applies Adam for gradient-based refinement.Experiments on three tasksthe multi-modal Rastrigin function, the CartPole reinforcement learning environment, and Support Vector Machine (SVM) hyperparameter tuning on a Public Benchmark Dataset -show significant gains over standalone GA and Adam.On the Rastrigin function, the objective value improved from 1.3 × 10⁻⁵ (GA) to nearly zero GA-Adam; in CartPole, the average reward increased from approximately 140 to approximately 490; and in SVM tuning, cross-validation accuracy reached 98.89%, demonstrating both improved convergence efficiency and more stable performance across runs.These results confirm the effectiveness of GA-Adam in achieving faster convergence and higher accuracy with moderate computational cost, making it suitable for hyperparameter tuning, control policy optimization, and multi-modal function optimization.