Bi-Directional Learning Particle Swarm Optimization for Large-Scale Optimization

Abstract

Large-scale optimization problems (LSOPs) are an essential research topic in the evolutionary computation (EC) community with two challenges: slow convergence in the huge search space and the trap in massive locally optimal solutions. To tackle these two challenges, this paper proposes a bi-directional learning particle swarm optimization (BLPSO) with two learning strategies, called diversity learning strategy (DLS) and convergence learning strategy (CLS). In DLS, we first propose a diversity evaluation mechanism based on locally sensitive hashing (LSH) to measure the diversity of individuals. Then the density individuals will learn from other dispersed individuals with good diversity to enhance the diversity of the population. In CLS, the inferior individuals will learn from other superior individuals with excellent evolution information to help the population accelerate convergence speed. These two learning strategies act in different roles and complement each other. With these two learning strategies, BLPSO achieves a balance between sufficient diversity and fast convergence in solving LSOPs. Two large-scale test suites, IEEE  CEC2010 and IEEE CEC2013, are used to test the performance between BLPSO and other state-of-the-art algorithms. The experimental results show that BLPSO outperforms other algorithms on both test suites, including the winner algorithms of the IEEE CEC2010 and IEEE CEC2012 competitions. Finally, BLPSO is applied to a large-scale portfolio optimization problem to illustrate its application capability.

Keywords: Bi-directional learning particle swarm optimization (BLPSO), large-scale optimization, particle swarm optimization, locally sensitive hashing (LSH)