Effects of repair procedures on the performance of EMO algorithms for multiobjective 0/1 knapsack problems

Multiobjective 0/1 knapsack problems have been used for examining the performance of EMO (evolutionary multiobjective optimization) algorithms in the literature. We demonstrate that their performance on such a test problem strongly depends on the choice of a repair procedure. We show through computational experiments that much better results are obtained from greedy repair based on a weighted scalar fitness function than the maximum profit/weight ratio, which has been often used for ordering items in many studies. This observation explains several reported results in comparative studies about the superiority of EMO algorithms with a weighted scalar fitness function. It is also shown that the performance of EMO algorithms based on Pareto ranking is significantly improved by the use of the weighted scalar fitness function in repair procedures. We also examine randomized greedy repair, where items are ordered based on the profit/weight ratio with respect to a randomly selected knapsack.