Directly Solving Multitasking Optimization Problems without Knowledge Transfer

In this work, we aim to solve multitasking optimization problems (MTOPs) without using knowledge transfer or other delicate strategies. We adopt a simple genetic framework (i.e., mating selection, offspring generation, and environmental selection) with novel variation operators, which separately solves each task with the total number of function evaluations being equally divided. That is, the optimization of each task is totally independent. We aim to achieve better performance than existing multitasking evolutionary algorithms (EAs) across all tasks, with the assistance of the versatile and learnable translation, scale, and rotation invariant variation operators. Through a series of comparative experiments on single- and multi-objective MTOPs, this work answers two questions: Can an EA without knowledge transfer outperform those with knowledge transfer on MTOPs? If so, can knowledge transfer further accelerate the convergence of such an EA?