A Dynamic Algorithm Configuration Framework Using Combinatorial Problem Features and Reinforcement Learning

This study explores the dynamic configuration of a population-based metaheuristic with reinforcement learning. Beyond achieving high performance, our dual focus involves utilizing hyperparameters as indicators for transitions between exploration and exploitation phases. We investigate how this information can be effectively harnessed for responsive balance tailored to each problem instance. Specifically, we analyze the potential of integrating the Local Optima Network (LON), an abstraction of the fitness landscape, to inform parameter generation. To study the relationship between indicators and responsive control, we embed the algorithm within a reinforcement learning framework.