Reinforcement Learning Applied to PSO for Multidimensional Knapsack Problem

The present work delves into the field of bio-inspired algorithms, focusing on enhancing the Particle Swarm Optimization (PSO) algorithm using QLearning. The main challenge these algorithms face is getting trapped in local optima, leading to inefficient use of computational resources. To mitigate this, we propose a solution that integrates QLearning with entropic diversity as a movement component in PSO. This integration aims to effectively detect and overcome stagnation in local optima. Our study conducts exhaustive tests on 30 instances of binary optimization problems from the OR-Library, specifically the Multidimensional Knapsack Problem. The results demonstrate a marginal but consistent improvement in the performance of the PSO-QL algorithm over standard PSO, particularly in instances with previously unknown solutions and in scenarios of more complex problems. Additionally, suggestions for future research directions are added, including comparing various movement techniques and adapting PSO to different forms of entropy, to further validate and refine the proposed approach.