Dynamic multitask optimization with improved knowledge transfer mechanism

Multitasking optimization (MTO) is promising to become the next-generation mainstream optimization paradigm for optimizing multiple tasks simultaneously with high efficiency and accuracy. However, despite dynamic tasks abound in the real world, such as flow shop scheduling, vehicle routing, IoT, machine learning, research on dynamic multitask optimization (DMTO) has been rarely reported. DMTO problems are more challenging than MTO with static tasks or a single dynamic optimization. In this paper, a dynamic multitask optimization algorithm with an improved knowledge transfer mechanism (IK_DMTO) is proposed to solve the DMTO problems. Firstly, an improved knowledge transfer mechanism is designed to promote knowledge utilization by conditionally selecting the scale of knowledge transfer and reduce negative migration by selectively performing the crossover operation between tasks. Secondly, a new individual information update strategy is applied to guide the individual updates, in which the leaders of the sub-populations formed during the knowledge transfer process are utilized to adjust the direction of individuals to make the utmost of knowledge between tasks, and an external archive management strategy is introduced to achieve a better distribution of non-dominated solutions. Finally, nine dynamic multi-objective multitask optimization (DMOMTO) problems are constructed with the dynamic multi-objective benchmark functions to verify the effectiveness of IK_DMTO. The experimental results show that IK_DMTO can perform well on convergence compared to the comparison algorithms.