A Learnable Competitive Swarm Operator Driven by Reinforcement Learning for Multi-Objective Optimization Problems

During the optimization process, producing high-quality individuals is always the goal of evolutionary algorithms. So, how can we generate such promising offspring individuals? The key lies in an optimizer with the potential to consistently produce superior individuals. However, designing the operator with learning ability in an optimizer has not been well explored in the existing research. To overcome this drawback, a learnable competition operator driven by reinforcement learning (RL) technique is proposed to generate high-quality individuals and further improve the performance of the algorithm. First, with the help of Q-learning, a classical RL technique, the size of elite particle set in the population is taken as the set of action, each elite particle corresponds to a action. The optimal action is considered to be the next state. Then, a three-dimensional Q-table is designed for the whole population, so that each particle can learn its elite particles, rather than randomly selecting elite particles to guide the particle's evolution process. Furthermore, a novel algorithm embedded with the learnable competition operator is proposed. Finally, compared with the other state-of-the-art algorithms, the experimental results demonstrate that the proposed algorithm significantly enhances the solution accuracy.