Improved Multi-Objective Binary Fish School for Feature Selection

The Multi-Objective Binary Fish School Search (MOBFSS) algorithm was proposed to solve optimization problems with two or three conflicting objectives and operating on discrete binary variables. The original proposal revealed good accuracy but it also exhibited a high computational cost. Here, we present strategies to obtain an improved version of MOBFSS that reaches lower Pareto fronts for minimization problems at a better computational cost. We also deploy local search procedures as proposed in BMOPSO-CDRLS to find solutions closer to the optimal solution. The achieved results outperform the state-of-art algorithms BMOPSO-CDR and BMOPSO-CDRLS in feature selection problems for hypervolume optimization. Hence, this paper contributes to the literature in Swarm Intelligence by introducing several algorithms that can be applied to improve feature selection in the context of classification programs.