A Comparative Analysis of Reinforcement Learning Algorithms in a Hybrid Learning and Optimization Framework

The integration of Machine Learning (ML) and Constrained Optimization (CO) represents a promising avenue for enhancing decision-making capabilities in complex, uncertain environments. This thesis empirically evaluates the UNIFY framework, a novel approach that synergistically combines the predictive power of ML with the strategic prowess of CO. Focusing on the Energy Management System and Set Multi-cover with stochastic coverages problems, in this work we carry out a comparative analysis on the performance, efficiency, and scalability of four RL algorithms - A2C, PPO, TD3, and SAC - within the UNIFY framework. Empirical results reveal the strengths and limitations of these algorithms, highlighting SAC's superior sample efficiency and the benefits of training on multiple instances for improved model generalization. These findings underscore the potential of integrating ML and CO through RL, offering valuable insights for the development of advanced decision-making systems in various real-world applications.