A Level-Based Learning Differential Evolution for Multitask Optimization

Evolutionary multitask optimization (EMTO) is a new research topic that solves several tasks simultaneously by conducting knowledge transfer and leveraging the connections between them. Although the current EMTO algorithms have made progress to a certain extent, common issues of complicated operation, slow convergence speed and negative transfer are still need to be solved. In fact, the above drawbacks can be overcome by a systematic way to utilize higher-level individuals with superior fitness values to instruct the evolutionary process of lower-level individuals with inferior fitness values. Therefore, this article introduces a level-based learning (LL) strategy and DE/current to best/1 to EMTO, proposing an algorithm with LL strategy based on the differential evolution algorithm (LLDE). In LLDE, the individuals are divided into distinct levels based on their fitness. When information transfer occurs, individuals from a higher level of the current individual in source population will guide the evolution of individuals from the population of target task. At the self-evolution stage, there is also a certain probability of performing LL strategy, while individuals from higher level of the target population will guide the individuals at lower level of the target population as well. In particular, DE/current to best/1 is adopted combined with LL strategy, not only accelerates the convergence speed by introducing the global optimal solution, but also simplifies the process of the algorithm. The experimental results of LLDE on CEC17 benchmark indicate that LLDE outperforms other compared EMTO algorithms.