Everyone is welcome to attend the MSc (Mathematical and Computational Sciences) thesis defence of presenter Bowen Xu on Friday, July 26, at 10 am in HSB 105.

Title of the Thesis:  “Deep Reinforcement Learning for Smart Restarts In Exploration-only Exploitation-only Metaheuristic Hybrids”

Metaheuristic algorithms excel in addressing challenging optimization problems but often face the issue of premature convergence, limiting their potential during extended optimization periods. This research aims to overcome this limitation by integrating Reinforcement Learning to implement intelligent restart mechanisms in metaheuristic processes. The objective is to enhance the algorithms' ability to explore and exploit the solution space more effectively, thereby improving performance in complex optimization scenarios.

The study starts with a review of current metaheuristic algorithms, highlighting the issue of premature convergence. It then explores Reinforcement Learning principles, particularly their decision-making capabilities, to optimize metaheuristic performance. A novel framework is proposed where Reinforcement Learning agents monitor the optimization process, identify stagnation phases, and initiate intelligent restarts. These restarts are strategically guided by the agents' learned policies, ensuring diversified search when necessary and focused exploration of promising regions.

Experiments on benchmark optimization problems demonstrate that integrating Reinforcement Learning significantly mitigates premature convergence, leading to superior solution quality and robust performance across various domains. This research not only addresses a critical limitation in metaheuristic optimization but also suggests new applications of Reinforcement Learning for enhancing algorithmic efficiency. The findings underscore the potential of intelligent restart mechanisms to transform optimization, enabling more effective and adaptive metaheuristic solutions.