Comparison of Multiobjective Memetic Algorithms on 0/1 Knapsack Problems

The paper compares two well-known multiobjective memetic algorithms through computational experiments on 0/1 knapsack problems. The two algorithms are MOGLS (multiple objective genetic local search) of Jaszkiewicz and M-PAES (memetic Pareto archived evolution strategy) of Knowles & Corne. It is shown that the MOGLS with a sophisticated repair algorithm based on the current weight vector in the scalar fitness function has much higher search ability than the M-PAES with a simple repair algorithm. When they use the same simple repair algorithm, the M-PAES performs better overall. It is also shown that the diversity of non-dominated solutions obtained by the MPAES is small in comparison with the MOGLS. For improving the performance of the M-PAES, we examine the use of the scalar fitness function with a random weight vector in the selection procedure of parent solutions