Q-LEARNING ALGORITHM OPTIMIZATION VIA PARTICLE SWARM OPTIMIZATION APPLIED TO THE K-SERVO PROBLEM

The k-server problem is an online combinatorial optimization problem in which each movement decision is irrevocable and made without information about future demands. Q-Learning is a promising alternative for this scenario, but its performance is highly sensitive to hyperparameter calibration. This paper proposes a hybrid approach that uses Particle Swarm Optimization (PSO) to tune Q-Learning hyperparameters, combined with Double Q-Learning and Experience Replay. The approach was evaluated on instances parameterized by the Gini Index, covering Multifocal and Migratory dynamics. The experiments show that PSO-based tuning reduces movement cost and decreases the coefficient of variation by up to 45\% compared to fixed configurations, indicating greater learning stability.