Beta distribution-based monogamous pairs genetic algorithm for knowledge transfer in many-task optimization

Evolutionary multitasking aims to solve multiple independent optimization tasks simultaneously by excavating a population’s implicit parallelism. However, negative knowledge transfer (i.e., the performance of the target domain deteriorates after receiving information from the source domain) slows down the evolution of a task, even in the presence of positive knowledge transfer. To address this issue, a many-task monogamous pairs genetic algorithm (MTMopGA) with dynamic knowledge and intensity selection strategies is proposed to encourage positive knowledge transfer while reducing negative transfer effects between tasks. Specifically, the transfer behavioural patterns exhibited in the reproductive outcomes (between the source and the target tasks) are accommodated in a statistical model by assuming a Beta distribution. We then combine the results with a scalable task similarity measure, which considers both the decision space, and the objective space to estimate online the strength of future transfers between tasks. In addition, multiple crossover strategies are introduced to improve the solution quality. The efficacy of the proposed framework is evaluated through an extensive benchmark suite, as well as a real-world search-based software engineering (SBSE) application. The results indicate that the proposed framework can achieve efficient dynamic transfer. Insightful information into the influence of knowledge transfer is also discussed from the lens of biology.