Machine Learning Optimization Techniques: A Survey, Classification, Challenges, and Future Research Issues

Abstract

Optimization approaches in machine learning (ML) are essential for training models to obtain high performance across numerous domains. The article provides a comprehensive overview of ML optimization strategies, emphasizing their classification, obstacles, and potential areas for further study. We proceed with studying the historical progression of optimization methods, emphasizing significant developments and their influence on contemporary algorithms. We analyse the present research to identify widespread optimization algorithms and their uses in supervised learning, unsupervised learning, and reinforcement learning. Various common optimization constraints, including non-convexity, scalability issues, convergence problems, and concerns about robustness and generalization, are also explored. We suggest future research should focus on scalability problems, innovative optimization techniques, domain knowledge integration, and improving interpretability. The present study aims to provide an in-depth review of ML optimization by combining insights from historical advancements, literature evaluations, and current issues to guide future research efforts.