Solving many-task optimization problems via online intertask learning

Evolutionary multitasking optimization (EMTO) is a newly emerging search paradigm for tackling multiple problems concurrently via integrating population-based meta-heuristic method and transfer learning tech-nique, which has attracted considerable interest in computational intelligence community. Notably, effective knowledge transfer across tasks is the main challenge for intertask synergies in EMTO. Particularly, how to select proper auxiliary tasks for each constitutive task, how to adapt the intensity of intertask knowledge transfer and how to narrow the discrepancy between tasks are three key concerns in complicated many-task scenario. However, these issues are rarely studied jointly till date. Bearing this in mind, this article introduces a novel adaptive EMTO solver for many-task transfer optimization. Specifically, we develop an adaptive task selection mechanism to choose sources reasonably by the aid of maximum mean discrepancy. Besides, multi-armed bandit model is employed to control the intensity of knowledge transfer across tasks. In addition, restricted Boltzmann machine is utilized to extract latent features between tasks so that their discrepancy is reduced. Experiments are conducted on a series of numerical benchmarks to investigate the performance of our proposal, the comparison results reveal that our proposal manages to solve EMTO problems competitively when compared to several existing counterparts.

Automated summary

Evolutionary multitasking optimization (EMTO) is a new search paradigm that tackles multiple problems concurrently. The main challenge in EMTO is effective knowledge transfer across tasks. This article introduces an adaptive EMTO solver that addresses the task selection, knowledge transfer intensity, and discrepancy reduction issues. Experimental results demonstrate competitive performance compared to existing counterparts.