A fuzzy system based self-adaptive memetic algorithm using population diversity control for evolutionary multi-objective optimization.

Simulated by nature's evolution, numerous evolutionary algorithms had been proposed. These algorithms perform better for a particular problem domain and extensive parameter fine tuning and adaptations are required in optimizing problems of varied domain. This paper aims to develop robust and self-adaptive memetic algorithm by combining Differential Evolution based algorithm, a popular population based global search method with the Controlled Local search procedure to solve multi-objective optimization problems. Memetic Algorithm is an enhanced evolutionary algorithm, it combines global search method with local search techniques for faster convergence. Memetic algorithm improves both exploration and exploitation, preventing premature convergence and also refines the current best solutions efficiently. Proposed algorithm is named as Fuzzy based Memetic Algorithm using Diversity control (F-MAD). In F-MAD, population diversity is controlled through the control parameters self-adaptation of Differential Evolution algorithm (DE) such as, crossover rate and scaling factor by using two fuzzy systems. A controlled local search procedure is adapted for guiding convergence process thus balancing explore-exploit cycle. The control parameter self-adaptation and enhanced selection method with controlled local search method aid population diversity control in decision space and attaining optimal solutions with uniform distribution in terms of diversity and convergence metrics in objective space. These characteristics help the proposed method suitable to be extended to different application domain without the need of trial-and-error fine tuning of the parameters. The performance is tested through standard benchmark test problems-CEC 2009 test problems and DTLZ test problem and further validated through performance metrics and statistical test. It is compared with popular optimization algorithms and experiment results indicate that F-MAD perform well than State of-The-Art (SOTA) algorithms taken for comparison. F-MAD algorithm attains better results for 8 out of 10 CEC 2009 test problems (UF1-UF10) when compared to 20 other algorithms taken for comparison. For DTLZ problems, F-MAD attains better results for ALL 7 problems (DTLZ 1-DTLZ7) when compared to 8 other SOTA algorithms. The performance is further evaluated using Friedman rank test and the proposed F-MAD significantly outperformed other algorithms.